A photovoltaic output prediction method and system based on VMD-feature migration fusion in a data-poor scenario

Abstract

The application discloses a photovoltaic output prediction method and system based on VMD-feature migration fusion in a poor data scene, and is applied to the photovoltaic output prediction field, and comprises the following steps: constructing a multi-dimensional feature set according to photovoltaic output time series data and meteorological time series data of a first station; decomposing the photovoltaic output time series data into three frequency band components; determining a prediction result correction rule according to a first low-frequency component; establishing a mapping relationship between the multi-dimensional feature set and a second middle-frequency component and determining model parameters; constructing a time series prediction model with meteorological data and a third high-frequency component as input; obtaining a basic prediction model from the mapping relationship, the time series prediction model and the correction rule; adjusting parameters of the basic model according to second station data to obtain an adaptive model; and inputting historical data before a to-be-predicted period into the adaptive model to obtain a prediction result. The application can significantly improve prediction accuracy under the condition that effective samples of a target station are insufficient, and meets the prediction demand in a poor data scene.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic power output prediction technology, and in particular to a photovoltaic power output prediction method and system based on VMD-feature transfer fusion in data-poor scenarios. Background Technology

[0002] As the global energy transition progresses, the proportion of photovoltaic (PV) power generation in the power system continues to increase. However, PV power output is characterized by significant intermittency, volatility, and randomness, and its uncertainty poses serious challenges to the safe and stable operation of the power grid, electricity spot market transactions, and coordinated dispatch of power generation, grid, load, and energy storage.

[0003] In existing technologies, high-precision prediction of photovoltaic (PV) power output largely relies on large-scale historical labeled data of the target domain for model training. However, when the target domain is a newly commissioned power plant, a newly built distributed PV system, or an area with incomplete data collection, the number of effective labeled samples is severely insufficient. Traditional data-driven models, when faced with insufficient samples, cannot fully learn the data distribution patterns, leading to a significant decrease in prediction accuracy and a marked deterioration in model generalization ability, making it difficult to meet the demand for high-precision PV power output prediction in data-scarce scenarios. Summary of the Invention

[0004] This invention provides a photovoltaic power output prediction method and system based on VMD-feature transfer fusion in data-scarce scenarios, to solve the technical problem that it is difficult to build a high-precision photovoltaic power output prediction model for target sites when the effective labeled sample size is insufficient. The invention enables the basic prediction model built from the source site to obtain high-precision and high-generalization photovoltaic power output prediction results after being adapted with a small amount of target site data.

[0005] To address the aforementioned technical problems, embodiments of the present invention provide a photovoltaic power output prediction method based on VMD-feature transfer fusion in data-scarce scenarios, comprising: At least based on the photovoltaic power output time-series data of the first site and the corresponding meteorological time-series data, a multi-dimensional feature set is constructed; Based on the frequency band characteristics of the photovoltaic power output time series data, the photovoltaic power output time series data is decomposed into a first low-frequency component, a second medium-frequency component, and a third high-frequency component; Based on the first low-frequency component, determine the prediction result correction rules of the basic prediction model; Establish a mapping relationship between the multidimensional feature set and the second intermediate frequency component, and determine the model parameters of the mapping relationship through adaptive optimization; Using the meteorological time-series data and the third high-frequency component as input, a time-series prediction model corresponding to the third high-frequency component is constructed according to the model parameters; The basic prediction model is obtained from the mapping relationship, the time series prediction model, and the prediction result correction rules; Based on the photovoltaic power output time series data of the second site and the corresponding meteorological time series data, the parameters of the basic prediction model are adjusted to obtain an adapted photovoltaic power output prediction model. The historical time-series data of photovoltaic power output of the second station before the predicted period and the corresponding historical time-series data of meteorology are input into the adapted photovoltaic power output prediction model to obtain the photovoltaic power output prediction result.

[0006] As one preferred embodiment, the construction of a multi-dimensional feature set based at least on the photovoltaic output time-series data of the first site and the corresponding meteorological time-series data includes: Obtain the meteorological time-series data and time characteristic data of the first station; Multiple meteorological factors in the meteorological time-series data are combined and transformed to obtain derived meteorological characteristics; Lag characteristics are extracted from the photovoltaic power output time series data and the meteorological time series data to obtain lag time series features; The time feature data is converted into periodic encoding to obtain periodic encoded features; The multidimensional feature set is obtained by integrating the derived meteorological features, the lag time series features, and the periodic coding features.

[0007] As one preferred embodiment, the photovoltaic power output time-series data is decomposed, including: Variational mode decomposition is used to decompose the photovoltaic power output time series data.

[0008] As one preferred embodiment, the step of determining the prediction result correction rule of the basic prediction model based on the first low-frequency component includes: The first low-frequency component is used as the daily cycle baseline trend of photovoltaic power output; Based on the daily cycle baseline trend, physical constraints are set for the photovoltaic power output prediction results to obtain the prediction result correction rules.

[0009] As one preferred embodiment, establishing the mapping relationship between the multidimensional feature set and the second intermediate frequency component includes: The photovoltaic power output time series data and meteorological time series data of the first site are input into a long short-term memory network to obtain deep time series features; The deep temporal features are fused with the multidimensional feature set to obtain a fused feature set; Using the fused feature set as input and the second intermediate frequency component as the prediction target, the mapping relationship between the multidimensional feature set and the second intermediate frequency component is established.

[0010] As one preferred approach, the adaptive optimization is implemented using the particle swarm optimization algorithm.

[0011] As one preferred embodiment, the step of constructing a time-series prediction model corresponding to the third high-frequency component based on the meteorological time-series data and the third high-frequency component as input, according to the model parameters, includes: Configure the single-layer recurrent neural network structure of the time series prediction model according to the model parameters; Using the meteorological time-series data and the third high-frequency component as input, and combining the single-layer recurrent neural network structure, the time-series prediction model is constructed.

[0012] As one preferred embodiment, adjusting the parameters of the basic prediction model to obtain an adapted photovoltaic output prediction model includes: Obtain the photovoltaic power output time-series data and the corresponding meteorological time-series data of the second site; Based on the same frequency band characteristics as the first site, the photovoltaic output time series data of the second site is decomposed into corresponding low-frequency components, mid-frequency components and high-frequency components; The model parameters of the mapping relationship are adjusted according to the mid-frequency component, and the model parameters of the time-series prediction model are adjusted according to the high-frequency component to obtain the adapted photovoltaic power output prediction model.

[0013] As one preferred embodiment, the parameter adjustment of the basic prediction model further includes: Based on the ratio of the sample size at the second site to the sample size at the first site, select the parameter adjustment method; When the ratio is lower than a preset threshold, a freeze adjustment method is used to fix the bottom network parameters and update the top network parameters. When the ratio is not lower than a preset threshold, a full network adjustment method is adopted, and all network parameters are updated with the parameters of the basic prediction model as the initial values.

[0014] Another embodiment of the present invention provides a photovoltaic power output prediction system based on VMD-feature transfer fusion in a data-scarce scenario, comprising: The feature construction module is used to construct a multi-dimensional feature set based at least on the photovoltaic power output time-series data of the first site and the corresponding meteorological time-series data; The frequency band decomposition module is used to decompose the photovoltaic power output time series data into a first low-frequency component, a second medium-frequency component, and a third high-frequency component according to the frequency band characteristics of the photovoltaic power output time series data. The correction rule determination module is used to determine the correction rule for the prediction result of the basic prediction model based on the first low-frequency component. The mapping relationship construction module is used to establish the mapping relationship between the multidimensional feature set and the second intermediate frequency component, and to determine the model parameters of the mapping relationship through adaptive optimization; The time series prediction modeling module is used to construct a time series prediction model corresponding to the third high frequency component based on the meteorological time series data and the third high frequency component as input and the model parameters. The basic model generation module is used to obtain the basic prediction model from the mapping relationship, the time series prediction model, and the prediction result correction rules; The model adaptation module is used to adjust the parameters of the basic prediction model based on the photovoltaic power output time series data of the second site and the corresponding meteorological time series data to obtain an adapted photovoltaic power output prediction model. The prediction execution module is used to input the historical time series data of photovoltaic power output of the second site before the period to be predicted and the corresponding historical time series data of meteorology into the adapted photovoltaic power output prediction model to obtain the photovoltaic power output prediction result.

[0015] Compared with the prior art, the beneficial effects of the embodiments of the present invention are at least one of the following: (1) This invention constructs a multi-dimensional feature set based on at least the photovoltaic power output time-series data and meteorological time-series data of the first site; decomposes the photovoltaic power output time-series data into a first low-frequency component, a second mid-frequency component, and a third high-frequency component according to frequency band characteristics; then determines the prediction result correction rules based on the first low-frequency component and establishes the mapping relationship between the multi-dimensional feature set and the second mid-frequency component; constructs a time-series prediction model using meteorological time-series data and the third high-frequency component as inputs; and finally forms a basic prediction model together with the mapping relationship, the time-series prediction model, and the correction rules. This process decouples the regular trend, main fluctuation component, and random disturbance component of photovoltaic power output, enabling each frequency band component to adopt a modeling strategy that matches its physical characteristics, effectively reducing the interference of non-stationary signals on model training. On this basis, the parameters of the basic prediction model are adjusted according to the data of the second site to obtain an adapted model, and the historical data before the period to be predicted are input into the adapted model to obtain the prediction result. Thus, this invention can significantly improve the accuracy and stability of photovoltaic power output prediction even when the effective labeled sample size of the target site is insufficient.

[0016] (2) This invention adopts a basic model construction strategy that combines frequency band decomposition and adaptive optimization to transform the non-stationary photovoltaic power output time series signal into multiple relatively stationary frequency band components, thereby reducing the model's dependence on massive training data. At the same time, by determining the model parameters with mapping relationships through adaptive optimization, the subjectivity and limitations of manual experience-based parameter tuning are avoided, enabling the model to have optimal prediction performance at the source site, and providing a high-quality source domain knowledge foundation for subsequent transfer learning. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating a photovoltaic power output prediction method based on VMD-feature transfer fusion in a data-poor scenario, according to one embodiment of the present invention. Figure 2 This is a schematic diagram of a photovoltaic power output prediction system based on VMD-feature transfer fusion in a data-poor scenario, according to one embodiment of the present invention.

[0018] Figure label: The modules are: 11. Feature construction module; 12. Frequency band decomposition module; 13. Correction rule determination module; 14. Mapping relationship construction module; 15. Time series prediction modeling module; 16. Basic model generation module; 17. Model adaptation module; and 18. Prediction execution module. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The purpose of providing these embodiments is to make the disclosure of the present invention more thorough and comprehensive. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0020] In the description of this application, the terms "first," "second," "third," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined with "first," "second," "third," etc., may explicitly or implicitly include one or more of that feature. In the description of this application, unless otherwise stated, "a plurality of" means two or more.

[0021] In the description of this application, it should be noted that, unless otherwise expressly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to fixed connections, detachable connections, or integral connections; they can refer to mechanical connections or electrical connections; they can refer to direct connections or indirect connections through an intermediate medium; and they can refer to the internal communication between two components. The terms "vertical," "horizontal," "left," "right," "upper," "lower," and similar expressions used herein are for illustrative purposes only and do not indicate or imply that the device or component referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as limiting the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items. Those skilled in the art can understand the specific meaning of the above terms in this application based on the specific circumstances.

[0022] In the description of this application, it should be noted that, unless otherwise defined, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art. The terminology used in this specification is for the purpose of describing specific embodiments only and is not intended to limit the invention. Those skilled in the art can understand the specific meaning of the above terms in this application based on the specific circumstances.

[0023] One embodiment of the present invention provides a photovoltaic power output prediction method based on VMD-feature transfer fusion in a data-scarce scenario. For details, please refer to [link to relevant documentation]. Figure 1 , Figure 1 The diagram shown illustrates a flowchart of a photovoltaic power output prediction method for a data-poor scenario in one embodiment of the present invention, which includes steps S1 to S8: S1: Construct a multi-dimensional feature set based at least on the photovoltaic power output time-series data of the first site and the corresponding meteorological time-series data; Photovoltaic power output is comprehensively affected by various meteorological factors such as solar radiation, temperature, and humidity, and exhibits significant time-series lag effects and daily, weekly, and annual cycle patterns. If only the original meteorological sequence is used as model input, the model will struggle to capture the coupled nonlinear relationships between multiple meteorological factors and the long-term and short-term dependencies in the time-series dimension. Therefore, step S1 constructs multi-dimensional features from the original data to provide the model with richer input information, thus compensating for the insufficient sample size in data-scarce scenarios.

[0024] The first site refers to a photovoltaic power station with sufficient historical data, used to provide source domain data for training the basic prediction model. Photovoltaic power output time-series data refers to the active power output values ​​of a photovoltaic power station at consecutive time points, arranged chronologically, with the unit being watts. Meteorological time-series data refers to the sequence of environmental factor observations that are strictly aligned with the photovoltaic power output time-series data in time, including three core factors: temperature, humidity, and solar radiation.

[0025] In this embodiment, the first site uses hourly measured data from a centralized photovoltaic power station for three consecutive years, with a total sample size of 26,280 data points, covering complete seasonal and meteorological cycle characteristics. The second site uses hourly measured data from a newly built distributed photovoltaic power station in the same climate zone, with a total sample size of 744 data points.

[0026] In this invention, "poor data" refers to application scenarios where the effective labeled sample size of the target site (second site) is less than 30% of the sample size of the source site (first site), covering typical engineering situations such as newly built photovoltaic power plants, distributed photovoltaic power plants, and old power plants with incomplete data acquisition systems.

[0027] A multidimensional feature set refers to a set of multidimensional input features extracted from raw time-series data through feature engineering, used to replace the raw data as input to a prediction model. This set is not a simple stacking of raw data, but a high-dimensional feature space constructed through various methods such as combination transformation, lag extraction, periodic encoding, and smoothing filtering, aiming to fully characterize the complex mapping relationship between meteorological factors and photovoltaic output.

[0028] Preferably, in one embodiment of the present invention, a multi-dimensional feature set is constructed based at least on the photovoltaic power output time-series data of the first site and the corresponding meteorological time-series data, including: Obtain meteorological time-series data and temporal characteristic data for the first station; Multiple meteorological factors in meteorological time-series data are combined and transformed to obtain derived meteorological characteristics; Lag characteristics were extracted from photovoltaic power output time series data and meteorological time series data to obtain lag time series features; Periodic encoding transformation is performed on time feature data to obtain periodic encoded features; By integrating derived meteorological features, lagging time series features, and periodic coding features, a multidimensional feature set is obtained.

[0029] Specifically, the system acquires photovoltaic (PV) output time-series data, meteorological time-series data, and time characteristic data from the first site. All data is time-resolutiond to the hour and aligned by timestamps. PV output time-series data is directly read from the grid-connected electricity meter or the Supervisory Control and Data Acquisition (SCADA) system at the PV power station, with one data point collected per hour. Meteorological time-series data is acquired from a weather station deployed at the PV power station site. The weather station is equipped with temperature, humidity, and solar radiation sensors, all synchronized with the power station's clock system to ensure precise time alignment between meteorological and PV output time-series data. Simultaneously, time characteristic data is acquired, including the hour (0-23), day of the week (1-7), month (1-12), and holiday identifiers for that time point. Holiday identifiers are set according to national statutory holidays, with statutory holidays marked as 1 and weekdays marked as 0.

[0030] After obtaining the raw data, it needs to undergo a three-stage preprocessing: (1) Outlier detection and replacement: The mean plus or minus 2.5 times the standard deviation is used as the outlier judgment threshold. Data points that exceed this range are identified as outliers and replaced with the median of the time series data of the same dimension.

[0031] (2) Missing value imputation: For missing values ​​with a continuous missing duration of less than 3 hours, linear interpolation is used to imput them. For missing values ​​with a continuous missing duration of 3 hours or more, the historical average of the same dimension in the same period of the past 7 days is used to imput them.

[0032] (3) Data smoothing: The two non-core feature columns of temperature and humidity are smoothed and denoised by a moving average method with a window size of 5. The two core feature columns of photovoltaic power output and solar radiation are retained without smoothing operation, so as to fully preserve the core physical information that affects photovoltaic power output.

[0033] After data preprocessing, a multidimensional feature set is constructed. First, multiple meteorological factors in the meteorological time-series data are combined and transformed to obtain derived meteorological features. These derived meteorological features are new features constructed by multiplicative combination or nonlinear transformation of the original meteorological factors, used to characterize complex relationships that cannot be expressed by a single meteorological factor.

[0034] Specifically, combinatorial transformations include two types of operations: (1) Constructing interactive features between meteorological factors, that is, combining two meteorological factors into new features through multiplication operations to characterize the synergistic effect of the two factors on photovoltaic power output. For example, the effect of solar radiation on power output differs under high-temperature conditions compared to non-high-temperature conditions, and the interactive features of solar radiation and temperature can capture this synergistic effect. The specific interactive features constructed include: the interaction between solar radiation and photovoltaic power output, the interaction between solar radiation and temperature, the interaction between humidity and solar radiation, the interaction between temperature and photovoltaic power output, and the interaction between humidity and temperature.

[0035] (2) Introduce square and cubic terms to solar radiation to characterize the nonlinear excitation effect of solar radiation on photovoltaic power output. This is because the relationship between solar radiation and photovoltaic power output is not strictly linear; power output increases more slowly when radiation is low and more rapidly when radiation is adequate.

[0036] Lag features were extracted from photovoltaic (PV) power output time-series data and meteorological time-series data to obtain lag time-series features. These lag time-series features use historical variable values ​​as input features for the current time, capturing the autocorrelation of time-series data across different time intervals and the delayed impact of meteorological factors on power output. Taking PV power output as an example, the 24-hour lag feature at time t is the PV power output value at time t-24. This feature reflects the autocorrelation of PV power output on a daily scale, meaning that the power output at this time today is strongly correlated with the power output at the same time yesterday.

[0037] Specifically, it is necessary to construct lag features for 1 hour, 2 hours, 3 hours, 6 hours, 12 hours, and 24 hours for the four core dimensions of solar radiation, photovoltaic output, temperature, and humidity. The lag features at different scales are used to capture dependencies over different time spans: 1-3 hour lags capture short-term fluctuations and continuity, 6-12 hour lags capture semi-daily cycle patterns, and 24-hour lags capture daily cycle patterns.

[0038] Periodic encoding transformation is performed on time feature data to obtain periodic encoded features. Periodic encoded features refer to the transformation of time variables (such as hours, days of the week, and months) that are periodic but have discontinuous numerical boundaries into numerical features that remain continuous at the boundaries of the cycle, so that the model can correctly understand the physical proximity of time variables at the beginning and end of the cycle.

[0039] Considering the inherent periodicity of time variables such as hours, days of the week, and months, directly using the original values ​​would disrupt the continuity at the boundaries of the period. For example, 11 PM and midnight differ by 23 numerically, but physically only by 1 hour. Directly using the original values ​​would cause the model to incorrectly assume a large difference between 11 PM and midnight. Sine-cosine encoding solves this problem by mapping periodic variables to points on a two-dimensional plane. Taking hours as an example, the hour value is converted into an angle (hour / 24 × 2π), and its sine and cosine values ​​are taken as two new features. In this way, 11 PM and midnight are very close in the sine-cosine space, consistent with their physical adjacency.

[0040] Similarly, the weekday variable is encoded with a 7-period cycle, and the month variable with a 12-period cycle. Furthermore, square and cube terms for the hour are introduced to enhance the model's ability to fit nonlinear changes within the daily cycle. Moving average features with a 3-hour sliding window are constructed for the three dimensions of solar radiation, photovoltaic output, and temperature to smooth short-term random fluctuations and highlight the core trends in output.

[0041] Finally, the aforementioned derived meteorological features, lagged time series features, periodic coding features, and time series smoothing features are horizontally spliced ​​along the feature dimension to form a unified multidimensional feature set.

[0042] S2: According to the frequency band characteristics of the photovoltaic power output time series data, the photovoltaic power output time series data is decomposed into a first low-frequency component, a second medium-frequency component and a third high-frequency component; In this embodiment, frequency band characteristics refer to the energy distribution characteristics of a signal in different frequency ranges. Photovoltaic power output time series data is essentially a non-stationary signal, and its fluctuation components exhibit differentiated distributions in the frequency domain: long-term trend changes are concentrated in the low-frequency band, mainly affected by regular factors such as solar altitude angle and seasonal changes; short-term fluctuations are concentrated in the mid-frequency band, mainly driven by meteorological factors such as cloud cover and temperature changes; random noise is distributed in the high-frequency band, mainly originating from measurement errors and micro-meteorological disturbances.

[0043] If the original signal is modeled directly, the mutual coupling of the frequency band components will increase the learning difficulty of the model. Therefore, step S2 decouples the original signal into three independent subsequences through frequency band decomposition, so that each component has more stable statistical characteristics and reduces the complexity of subsequent modeling.

[0044] In this embodiment, the photovoltaic power output time series data is decomposed by variational mode decomposition (VMD), which is an adaptive, non-recursive signal decomposition method that can decompose the original signal into several eigenmode functions with sparse characteristics, each mode being concentrated near its own center frequency.

[0045] Variational mode decomposition (VMD) is used instead of other signal decomposition methods because photovoltaic power output time-series data is non-stationary and nonlinear, and the actual acquired data often contains noise and outliers. Other decomposition methods, such as empirical mode decomposition (EMD), use a recursive screening approach to decompose layer by layer, which is sensitive to noise and sampling, and is prone to mode aliasing, i.e., components from different frequency bands are mixed into the same mode, resulting in unclear physical meanings of the decomposed components. Variational mode decomposition transforms the decomposition problem into a variational optimization problem, determining all modes at once by solving a constrained variational model. Each mode exhibits sparsity and compactness in the frequency domain, effectively suppressing mode aliasing. Moreover, variational mode decomposition has better robustness to noise and outliers, the decomposition results are stable, and the center frequencies of each component can be adaptively determined without manually setting the frequency band range.

[0046] Specifically, variational mode decomposition (VMD) algorithm is used to perform frequency band decomposition on the preprocessed photovoltaic power output time series data. The number of decomposition modes is set to 3, the penalty coefficient is 2500, the convergence accuracy is 1e-7, the DC component constraint is DC=0, and the initialization method is init=1. After decomposition, three subsequences are obtained, which are labeled as the first low-frequency component, the second mid-frequency component, and the third high-frequency component in order of increasing center frequency.

[0047] The first low-frequency component corresponds to the daily baseline trend of photovoltaic (PV) output, exhibiting gentle fluctuations and strong regularity. It reflects the control effect of macroscopic factors such as solar altitude angle and seasonal changes on output, and will be subsequently constrained and corrected using physical rules. The second mid-frequency component corresponds to the core fluctuation component of PV output, accounting for the highest energy proportion. It reflects the dominant influence of meteorological factors such as cloud cover and temperature changes on output, and will require a high-precision mapping model for fitting. The third high-frequency component corresponds to the random disturbance component, accounting for a relatively low energy proportion (usually less than 10% of the total). It characterizes rapid changes caused by random factors such as short-term cloud fluctuations and measurement noise, and will be fitted using a lightweight model.

[0048] After decomposition, maximum-min normalization is performed on the three components respectively, and the numerical range of each component is uniformly scaled to the [0,1] interval. The maximum and minimum values ​​of each component are saved for subsequent inverse normalization. The normalization formula is: the normalized value is equal to the original value minus the minimum value and divided by the difference between the maximum and minimum values.

[0049] S3: Based on the first low-frequency component, determine the prediction result correction rules of the basic prediction model; In this embodiment, the first low-frequency component is the daily periodic baseline trend component separated from the photovoltaic power output time-series data through variational mode decomposition. This component reflects the regular fluctuation pattern of photovoltaic power output throughout the day as the solar altitude angle changes. Its characteristics are smooth changes, significant periodicity, and strong predictability. The low-frequency component essentially represents the ideal power output curve of photovoltaic power output under the condition of no random disturbance. Therefore, in this embodiment, it can be used as a benchmark reference to judge whether the actual power output is reasonable.

[0050] A common problem in photovoltaic (PV) power output prediction is that purely data-driven models can still output non-zero predicted values ​​under physically unreasonable conditions. For example, during the night when solar radiation is zero, PV modules cannot generate electricity, and the model output should be strictly zero; during the daytime when solar radiation is extremely low, the output should also be close to zero. Traditional models, lacking physical constraints, often output small non-zero values ​​during these periods, causing a "zero-point drift" phenomenon that affects the engineering usability of the prediction results. In this embodiment, the correction rule is determined based on the first low-frequency component to solve this problem.

[0051] Preferably, in one embodiment of the present invention, determining the prediction result correction rule of the basic prediction model based on the first low-frequency component includes: The first low-frequency component is used as the daily cycle baseline trend of photovoltaic power output; Based on the daily cycle baseline trend, physical constraints are set for the photovoltaic power output prediction results, and the prediction result correction rules are obtained.

[0052] Specifically, the first low-frequency component is used as the daily cycle baseline trend of photovoltaic power output. The daily cycle baseline trend refers to the regular fluctuation curve of photovoltaic power output as the solar altitude angle changes throughout the day. This curve approaches zero at night, gradually rises after sunrise, reaches its peak around noon, and returns to zero after sunset.

[0053] Based on the daily cycle baseline trend, physical constraints are set for the photovoltaic power output prediction results. The core of the physical constraints is to determine whether the current moment meets the criterion of zero photovoltaic power output. In this embodiment, the criterion is divided into two categories: (1) The measured solar radiation at the current moment is less than or equal to zero, and the current moment is during the nighttime period. The nighttime period is defined as 18:00 on the current day to 6:00 on the next day. There is no solar radiation during this period, and the photovoltaic modules cannot generate electricity, so the predicted value must be zero.

[0054] (2) The current time is during daytime (6:00 to 18:00), but the measured value of solar radiation is less than or equal to zero. This usually occurs on cloudy or rainy days or before sunrise or after sunset when there is no effective radiation. At this time, the photovoltaic output should also be zero.

[0055] If either of the two judgment conditions mentioned above is met, a forced zeroing operation is triggered, and the photovoltaic output prediction value at that moment is modified to zero; if the zeroing condition is not met, the original prediction value of the model output remains unchanged. The above rules are the prediction result correction rules.

[0056] This rule is executed during the model prediction and inference phase. That is, after the model outputs the initial photovoltaic power output prediction result, the prediction result at each moment is verified according to the above judgment conditions. For prediction values ​​that meet the zeroing condition, a forced zeroing correction is performed, and for prediction values ​​that do not meet the condition, no correction is performed.

[0057] S4: Establish the mapping relationship between the multidimensional feature set and the second intermediate frequency component, and determine the model parameters of the mapping relationship through adaptive optimization; In this embodiment, the second intermediate frequency component is the main fluctuation component obtained through variational mode decomposition. It has the highest energy proportion and reflects the dominant influence of meteorological factors such as cloud cover and temperature changes on photovoltaic power output. It is the key component that the prediction model needs to fit. The multidimensional feature set is a high-dimensional set of input features extracted from the original data, including derived meteorological features, lagged time-series features, and periodic coding features. Step S4 establishes a nonlinear mapping relationship between the two to learn the complex correspondence between meteorological factors and the core fluctuations of photovoltaic power output.

[0058] The mapping relationship is established using a hybrid modeling strategy that combines Long Short-Term Memory (LSTM) networks with Random Forest (RF).

[0059] Long Short-Term Memory (LSTM) networks are a variant of recurrent neural networks (RNNs) that excels at processing time-series data. By introducing gating mechanisms (forget gate, input gate, output gate), they effectively solve the gradient vanishing and gradient exploding problems of traditional RNNs during long-sequence training, enabling them to capture long-short-term dependencies in time-series data. Random Forest is a machine learning method based on decision tree ensembles. By constructing multiple decision trees and averaging or voting on their predictions, it has advantages such as strong ability to handle high-dimensional features, good resistance to overfitting, and high interpretability. Combining the two allows us to leverage both the deep time-series features extracted by LSM networks and the advantages of random forests in multi-dimensional feature fusion modeling.

[0060] Preferably, in one embodiment of the present invention, establishing the mapping relationship between the multidimensional feature set and the second intermediate frequency component includes: The photovoltaic power output time series data and meteorological time series data of the first site were input into the long short-term memory network to obtain deep time series features; Deep temporal features are fused with multidimensional feature sets to obtain a fused feature set; Using the fused feature set as input and the second intermediate frequency component as the prediction target, a mapping relationship between the multidimensional feature set and the second intermediate frequency component is established.

[0061] Specifically, the preprocessed photovoltaic power output time-series data and meteorological time-series data are divided into sliding window segments according to a set time step. In this embodiment, the time step is set to 24 hours, that is, the photovoltaic power output at the current moment is predicted using historical data from the previous 24 hours. By traversing the normalized time-series data through the sliding window, a time-series input matrix with dimensions of [number of samples, 24, number of features] is constructed, where the number of features includes original features such as photovoltaic power output, temperature, humidity, and solar radiation, as well as their derived features; at the same time, a corresponding target label sequence is constructed, that is, the value of the second intermediate frequency component at the current moment.

[0062] The aforementioned temporal input matrix is ​​then fed into a Long Short-Term Memory (LSTM) network to obtain deep temporal features. In this embodiment, the LSM network adopts a bidirectional multi-layer structure: the input layer receives the temporal input matrix, which is then connected to the first bidirectional LSM layer. This layer processes temporal data from both forward and backward directions, capturing both past and future contextual information simultaneously. Next, a dropout layer is connected, randomly discarding some neuron outputs with a certain probability to suppress overfitting. A second bidirectional LSM layer and the dropout layer are then connected to further extract higher-order temporal features. Finally, a third bidirectional LSM layer and a fully connected layer are connected to map the temporal features to a specified dimension.

[0063] In this embodiment, the dropout rate of the dropout layer is set to 0.2, the output dimension of the fully connected layer is 128, and the hidden layer dimensions of the three bidirectional long short-term memory layers are all determined by particle swarm optimization (PSO) algorithm, with an optimization range of 128 to 256. The model uses the Adam optimizer, and the loss function is mean squared error (MSE). An early stopping mechanism is enabled during training, terminating training when the validation set loss does not decrease for 15 consecutive rounds. The 128-dimensional output of the fully connected layer after training is used as the extracted deep temporal features.

[0064] Deep temporal features are fused with a multidimensional feature set to obtain a fused feature set, which includes derived meteorological features, lagged temporal features, periodic coding features, and temporal smoothing features. In this embodiment, the fusion operation adopts a horizontal concatenation method, directly concatenating the 128-dimensional vector of the deep temporal features with the multidimensional vector of the multidimensional feature set along the feature dimensions to form a unified high-dimensional fused feature set.

[0065] Using the fused feature set as input and the second intermediate frequency component as the prediction target, a random forest regressor is trained to establish a mapping relationship. The core parameters of the random forest regressor, including the number of decision trees and the maximum depth, are both determined using particle swarm optimization. The number of decision trees is optimized from 100 to 200, and the maximum depth from 10 to 20. Parallel computing is employed during training to improve efficiency, and the model's fit is validated using out-of-bag (OOB) data. After training, the random forest regressor establishes a nonlinear mapping relationship from the fused feature set to the second intermediate frequency component.

[0066] It should be noted that, in this embodiment, the model parameters in the above mapping relationship (including the hidden layer dimension, stacking layer number, learning rate of the Long Short-Term Memory network, and the number of decision trees and maximum depth of the Random Forest) are all determined by an adaptive optimization method, which is implemented using the particle swarm optimization algorithm.

[0067] Particle swarm optimization (PSO) simulates the foraging behavior of bird flocks, treating each candidate hyperparameter combination as a particle in the search space. Each particle dynamically updates its position and velocity based on its own historical best position and the global best position of the group.

[0068] Specifically, the particle swarm is first initialized, with each particle's position vector containing the hyperparameter values ​​to be optimized. The mean squared error of predictions on the validation set is used as the fitness value to evaluate the performance of each particle. The particle velocity and position are iteratively updated until the maximum number of iterations is reached or the fitness value converges. Finally, the globally optimal particle position is output as the model parameters. In this embodiment, the number of particles is set to 15, the maximum number of iterations is set to 15, the inertia weight is set to 0.8, and both the cognitive coefficient and the social coefficient are set to 0.5. Through this adaptive optimization process, the optimal combination of hyperparameters for model prediction performance can be found in the global search space.

[0069] S5: Using the meteorological time series data and the third high-frequency component as input, construct the time series prediction model corresponding to the third high-frequency component according to the model parameters; In this embodiment, the third high-frequency component is a high-frequency perturbation component obtained through variational mode decomposition, characterizing rapid changes caused by random factors such as short-term cloud fluctuations and measurement noise. This component is characterized by high fluctuation frequency, small amplitude, and strong randomness, and its energy proportion is usually less than 10% of the total photovoltaic output energy, lacking obvious long-term time-series dependence. If the same complex model (a fusion of multilayer long short-term memory network and random forest) as the mid-frequency component is used to model this component, the number of model parameters will be too large, which is very likely to cause overfitting in the case of scarce data. Therefore, step S5 adopts a lightweight modeling strategy.

[0070] Recurrent neural networks (RNNs) are a type of neural network structure specifically designed for processing time-series data. Their core characteristic is the presence of recurrent connections within the network, allowing information to be transferred between time steps, thus enabling the capture of temporal dependencies in sequential data. The time-series prediction model for the third high-frequency component shares the same set of model parameters, determined through adaptive optimization, with the mapping relationship between the high-frequency component and the mid-frequency component. This includes hyperparameter configurations such as the learning rate. This parameter-sharing mechanism reduces the number of parameters requiring additional optimization and ensures the model possesses reasonable basic performance during the source domain training phase.

[0071] Preferably, in one embodiment of the present invention, a time-series prediction model corresponding to the third high-frequency component is constructed based on meteorological time-series data and a third high-frequency component as input, according to model parameters, including: Configure the single-layer recurrent neural network structure of the time series prediction model according to the model parameters; Using meteorological time-series data and the third high-frequency component as input, a time-series prediction model is constructed by combining a single-layer recurrent neural network structure.

[0072] Specifically, a single-layer recurrent neural network structure for the time series prediction model is configured according to the model parameters. This single-layer recurrent neural network contains a recurrent layer that receives the current input and the hidden state passed from the previous time step at each time step, passing the hidden state to the next time step through recurrent connections. The hidden layer dimension is set lower than that of the mid-frequency component long short-term memory network; in this embodiment, it is set to 64. The learning rate is the optimal value determined by the particle swarm optimization algorithm; in this embodiment, it is set to 0.0005. No dropout layer is set, or the dropout rate is set to no more than 0.1, to retain the sensitivity to high-frequency random fluctuations. The activation function is the rectified linear unit (ReLU).

[0073] Using meteorological time-series data and the third high-frequency component as input, and combining them with a pre-configured single-layer recurrent neural network structure, a time-series prediction model is constructed.

[0074] The model's input data includes meteorological time-series data (including three core factors: temperature, humidity, and solar radiation) corresponding to the time to be predicted, and the 24-hour historical sequence of the third high-frequency component itself, i.e., the high-frequency component values ​​24 times before the time to be predicted. These two types of data are serialized and organized according to a time step of 24, forming an input tensor with dimensions [number of samples, 24, number of input features], where the number of input features includes the number of meteorological factors and the number of historical values ​​of the high-frequency component.

[0075] In this implementation, the time series prediction model takes the aforementioned input tensor as input and the third high-frequency component at the time to be predicted as the output target. During training, an optimizer (Adam) and a loss function (mean squared error) with the same mapping relationship as the intermediate frequency component are used, and the network weights are updated through the backpropagation algorithm. Due to its simple model structure and few parameters, the training convergence speed is fast, and it has good generalization performance under small sample conditions.

[0076] S6: The basic prediction model is obtained from the mapping relationship, the time series prediction model, and the prediction result correction rules; The mapping relationship, time-series prediction model, and prediction result correction rules obtained in the aforementioned steps, i.e., the three sub-models of the basic prediction model, correspond to the prediction tasks of the three frequency band components, respectively.

[0077] Specifically, the mapping relationship is used as the prediction sub-model for the mid-frequency component. This sub-model receives a multi-dimensional feature set as input, extracts deep temporal features through a long short-term memory network, and then fuses them with the multi-dimensional feature set before inputting into a random forest regressor to output the predicted value of the second mid-frequency component. The time-series prediction model is used as the prediction sub-model for the high-frequency component. This sub-model receives meteorological time-series data and the 24-hour historical sequence of the third high-frequency component as input, and outputs the predicted value of the third high-frequency component through a single-layer recurrent neural network. The prediction result correction rule is used as a post-processing module to perform physical constraint verification on the final prediction result. This module does not participate in model training but only executes during the prediction inference stage. Its input is the initial predicted value of photovoltaic power output after superposition and reconstruction, and its output is the corrected predicted value of photovoltaic power output.

[0078] The three components are encapsulated and integrated to form a basic prediction model. Specifically, after the mid-frequency and high-frequency sub-models have converged during training, the weight parameters of the Long Short-Term Memory network, the model parameters of the Random Forest regressor, the weight parameters of the Recurrent Neural Network, and the normalization parameters of each component are stored together. During prediction, the three sub-models run independently, each outputting the predicted value for its corresponding frequency band component. The three predicted values ​​are linearly superimposed to obtain the initial photovoltaic power output prediction value. Finally, the prediction result correction rule is called to perform physical constraint verification and correction on the initial photovoltaic power output prediction value, and the final photovoltaic power output prediction result is output.

[0079] S7: Based on the photovoltaic power output time series data of the second site and the corresponding meteorological time series data, adjust the parameters of the basic prediction model to obtain an adapted photovoltaic power output prediction model; In this embodiment, the second site refers to the photovoltaic power station in the target domain, typically a newly built and operational power station, distributed photovoltaic, or an area with incomplete data collection. Its effective labeled sample size is small, belonging to a data-poor scenario. The basic prediction model is a prediction model trained based on the full data of the first site (source domain), and already possesses the basic modeling capability for photovoltaic output. Step S7 involves transferring the source domain knowledge to the target domain, adjusting the parameters of the basic prediction model using a small amount of target domain sample data to adapt it to the meteorological environment and output characteristics of the target domain, thereby generating a target domain-specific adapted photovoltaic output prediction model.

[0080] Because the first and second sites differ in meteorological conditions, geographical environment, and equipment parameters, parameter adjustments are necessary instead of directly using the basic prediction model. For example, the solar radiation intensity, temperature variation range, and cloud cover patterns may differ between the two sites, leading to discrepancies in the impact of the same meteorological factor on photovoltaic output. If the basic prediction model is directly used to predict the second site, the model cannot perceive these differences, resulting in decreased prediction accuracy. Fine-tuning the model parameters using a small number of target domain samples allows the model to quickly learn the data distribution characteristics of the target domain.

[0081] Among them, the prediction result correction rule corresponding to the low frequency component is no longer involved in parameter adjustment, because the rule is based on the objective physical law of photovoltaic power output (power output is zero when solar radiation is zero), and this law has universality across different sites.

[0082] Preferably, in one embodiment of the present invention, adjusting the parameters of the basic prediction model to obtain an adapted photovoltaic output prediction model includes: Obtain the photovoltaic power output time-series data and corresponding meteorological time-series data of the second site; Based on the same frequency band characteristics as the first site, the photovoltaic output time series data of the second site is decomposed into corresponding low-frequency components, mid-frequency components and high-frequency components; The model parameters of the mapping relationship are adjusted based on the intermediate frequency component, and the model parameters of the time series prediction model are adjusted based on the high frequency component to obtain an adapted photovoltaic power output prediction model.

[0083] Specifically, the photovoltaic (PV) output time-series data and corresponding meteorological time-series data for the second site are acquired. All data are time-resolutiond in hours and aligned by timestamps. The data acquisition method is the same as for the first site: PV output time-series data is read from the power station's electricity meters or data acquisition and monitoring control system, meteorological data is acquired from the on-site meteorological station, and time feature data is automatically generated based on the collection time point. Since the second site is in a data-poor scenario, its effective sample size is typically small, for example, it may only have one month or several weeks of historical data.

[0084] Based on the same frequency band characteristics as the first site, the photovoltaic output time series data of the second site is decomposed into corresponding low-frequency, mid-frequency, and high-frequency components. Specifically, the photovoltaic output time series data of the second site is decomposed using the same variational mode decomposition parameters as the first site (3 decomposition modes, 2500 penalty coefficient, and 1e-7 convergence accuracy) to obtain the low-frequency, mid-frequency, and high-frequency components of the second site.

[0085] The model parameters for the mapping relationship are adjusted based on the intermediate frequency components. Specifically, during the adjustment, the meteorological time-series data and multidimensional feature set of the second station are used as inputs, and the intermediate frequency components of the second station are used as the prediction targets. The parameters of the long short-term memory network and random forest regressor in the basic prediction model are updated. Depending on the sample size of the second station, either a freeze adjustment method or a full network adjustment method is selected.

[0086] Preferably, in one embodiment of the present invention, adjusting the parameters of the basic prediction model further includes: Based on the ratio of the sample size at the second site to the sample size at the first site, select the parameter adjustment method; When the ratio is lower than the preset threshold, the freeze adjustment method is used to fix the bottom network parameters and update the top network parameters; When the ratio is not lower than the preset threshold, the whole network adjustment method is adopted, and all network parameters are updated with the parameters of the basic prediction model as the initial values.

[0087] The freeze adjustment method is suitable for scenarios with extremely small sample sizes (e.g., the sample size of the second site is less than 10% of the sample size of the first site). In this case, the weight gradients of the bottom layer of the Long Short-Term Memory network are fixed, preserving the general temporal feature representation ability learned from the source domain. Only the parameters of the top fully connected layer and the output layer parameters of the random forest regressor are updated. This method can avoid overfitting of the entire network caused by small sample data.

[0088] The full network adjustment approach is suitable for scenarios where the sample size is relatively insufficient but still considerable (e.g., the sample size of the second site is 10% to 30% of that of the first site). In this case, the parameters of the base prediction model are used as the starting point for network initialization, with a learning rate lower than that used during the base model training phase (e.g., 0.0005). A full network backpropagation update is performed on the target domain dataset, simultaneously updating all layer parameters of the Long Short-Term Memory network and all decision tree parameters of the Random Forest regressor. This approach allows the model to better adapt to the data distribution of the target domain.

[0089] When the sample size of the second site is not less than 30% of the sample size of the first site, it does not fall under the poor data scenario defined in this invention, and the basic prediction model can be used directly for prediction.

[0090] The model parameters of the time-series forecasting model are adjusted based on the high-frequency components. Specifically, when adjusting the parameters, meteorological time-series data from the second station and historical sequences of the third high-frequency component are used as inputs, and the high-frequency component of the second station is used as the prediction target to update the parameters of the single-layer recurrent neural network in the basic forecasting model. The adjustment method is also selected based on the sample size, choosing between freeze adjustment or full network adjustment, and the specific strategy is consistent with the adjustment method for the mid-frequency component mapping relationship.

[0091] The adjusted mid-frequency component mapping relationship, the adjusted high-frequency component time-series prediction model, and the correction rules for the low-frequency component prediction results that were not adjusted are recombined to obtain the adapted photovoltaic power output prediction model. This adapted model inherits the overall architecture of the basic prediction model, and its parameters have been optimized for the data distribution of the second site, enabling high-precision prediction in the data-poor scenario of the target domain.

[0092] S8: Input the historical time series data of photovoltaic power output of the second station before the period to be predicted and the corresponding historical time series data of meteorology into the adapted photovoltaic power output prediction model to obtain the photovoltaic power output prediction result.

[0093] The period to be predicted refers to the future time period for which photovoltaic power output needs to be predicted, such as the next 24 hours. The historical time-series data of photovoltaic power output and the corresponding historical meteorological time-series data before the period to be predicted refer to historical observation data at a certain time step before the time to be predicted, which are used as input features for the prediction model.

[0094] Specifically, historical time-series data of photovoltaic (PV) output and corresponding meteorological data for the second site prior to the forecast period are obtained. The length of the historical data must meet the following requirements: for the mapping relationship of the mid-frequency component, the photovoltaic output time-series data and meteorological data (including temperature, humidity, and solar radiation) for the previous 24 hours are needed to construct the time-series input matrix; for the time-series prediction model of the high-frequency component, the meteorological data for the previous 24 hours and the historical sequence of the third high-frequency component are required. All data undergo the same preprocessing procedures as in the training phase, including outlier replacement, missing value imputation, and normalization.

[0095] The aforementioned historical data is then input into the adapted photovoltaic (PV) power output prediction model. Internally, the model first performs frequency band decomposition on the input data. It should be noted that during the prediction phase, the historical time-series PV power output data in the input data has already yielded corresponding frequency band components through variational mode decomposition. However, the power output at the time of prediction is unknown; therefore, the model only uses information from the historical data for forward inference. Specifically, the mid-frequency component mapping relationship in the model uses a multi-dimensional feature set constructed from historical meteorological data and historical PV power output time-series data as input, and outputs the predicted value of the second mid-frequency component; the high-frequency component time-series prediction model uses historical meteorological data and the historical sequence of the third high-frequency component as input, and outputs the predicted value of the third high-frequency component.

[0096] After obtaining the predicted values ​​of each frequency band component, the model performs a linear superposition and reconstruction operation. Specifically, the predicted values ​​of the second intermediate frequency component and the third high frequency component are added together time-by-time to obtain the initial photovoltaic power output prediction value.

[0097] After the superposition and reconstruction are completed, the model calls the prediction result correction rules to perform time-step physical constraint verification and correction on the initial photovoltaic power output prediction value. The correction rules read the measured solar radiation value at the current prediction time (if the measured solar radiation value at the time to be predicted is not available, the predicted radiation value provided by numerical weather prediction can be used) and time period information time by time. If the physical judgment condition of zero photovoltaic power output is met (nighttime and solar radiation is not greater than zero, or daytime and solar radiation is not greater than zero), a forced zeroing operation is triggered, and the photovoltaic power output prediction value at that time is modified to zero; if the zeroing condition is not met, the initial prediction value output by the model remains unchanged.

[0098] After correction, the prediction results are denormalized. Denormalization uses the maximum and minimum values ​​of each component saved during the training phase to restore the predicted values ​​from the normalized interval to the original physical dimension interval of photovoltaic output. The final photovoltaic output prediction results are output as a CSV data file containing hourly dimensions, actual photovoltaic output (if applicable), and predicted photovoltaic output. This file can be directly connected to the power dispatch system and the electricity spot market trading platform. Simultaneously, the system automatically calculates and outputs prediction accuracy evaluation indicators, including root mean square error, mean absolute percentage error, and coefficient of determination, and generates a visual chart comparing the actual and predicted photovoltaic output values ​​for the prediction period.

[0099] Another embodiment of the present invention provides a photovoltaic power output prediction system based on VMD-feature transfer fusion in a data-scarce scenario. For details, please refer to [link to relevant documentation]. Figure 2 , Figure 2 The diagram shown illustrates a photovoltaic power output prediction system in a data-poor scenario according to one embodiment of the present invention, which includes: Feature construction module 11 is used to construct a multi-dimensional feature set based at least on the photovoltaic power output time series data of the first site and the corresponding meteorological time series data; The frequency band decomposition module 12 is used to decompose the photovoltaic power output time series data into a first low-frequency component, a second medium-frequency component and a third high-frequency component according to the frequency band characteristics of the photovoltaic power output time series data. The correction rule determination module 13 is used to determine the correction rule for the prediction result of the basic prediction model based on the first low-frequency component. The mapping relationship construction module 14 is used to establish the mapping relationship between the multidimensional feature set and the second intermediate frequency component, and to determine the model parameters of the mapping relationship through adaptive optimization. The time series forecasting modeling module 15 is used to construct a time series forecasting model corresponding to the third high frequency component based on the model parameters, using meteorological time series data and the third high frequency component as input. The basic model generation module 16 is used to obtain the basic prediction model from the mapping relationship, the time series prediction model and the prediction result correction rules; Model adaptation module 17 is used to adjust the parameters of the basic prediction model based on the photovoltaic power output time series data of the second site and the corresponding meteorological time series data to obtain an adapted photovoltaic power output prediction model. The prediction execution module 18 is used to input the historical time series data of photovoltaic power output of the second station before the period to be predicted and the corresponding historical time series data of meteorology into the adapted photovoltaic power output prediction model to obtain the photovoltaic power output prediction result.

[0100] Compared with the prior art, the beneficial effects of the embodiments of the present invention are at least one of the following: (1) In the target site adaptation stage, this invention only requires a small amount of sample data to adjust the parameters of the basic prediction model, without having to train the model from scratch. This transfer learning mechanism significantly reduces the requirement for the amount of historical labeled data at the target site, enabling high-precision photovoltaic output prediction in data-poor scenarios such as newly built photovoltaic power plants, distributed photovoltaics, and areas with incomplete data collection, and significantly improving the generalization ability and feasibility of the model in engineering applications.

[0101] (2) This invention determines the prediction result correction rule based on the first low-frequency component and performs constraint correction on cases that do not conform to the physical laws of photovoltaic power generation before the prediction result is output. This mechanism eliminates the problem of non-zero prediction value anomalies that may occur during periods without solar radiation in traditional data-driven models, making the prediction result fully conform to the objective physical laws of photovoltaic power output. It can be directly connected to production links such as power dispatching and operation and power spot market transactions, and has extremely strong engineering practical value.

[0102] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A photovoltaic power output prediction method based on VMD-feature transfer fusion in a data-scarce scenario, characterized in that, include: At least based on the photovoltaic power output time-series data of the first site and the corresponding meteorological time-series data, a multi-dimensional feature set is constructed; Based on the frequency band characteristics of the photovoltaic power output time series data, the photovoltaic power output time series data is decomposed into a first low-frequency component, a second medium-frequency component, and a third high-frequency component; Based on the first low-frequency component, determine the prediction result correction rules of the basic prediction model; Establish a mapping relationship between the multidimensional feature set and the second intermediate frequency component, and determine the model parameters of the mapping relationship through adaptive optimization; Using the meteorological time-series data and the third high-frequency component as input, a time-series prediction model corresponding to the third high-frequency component is constructed according to the model parameters; The basic prediction model is obtained from the mapping relationship, the time series prediction model, and the prediction result correction rules; Based on the photovoltaic power output time series data of the second site and the corresponding meteorological time series data, the parameters of the basic prediction model are adjusted to obtain an adapted photovoltaic power output prediction model. The historical time series data of photovoltaic power output of the second station before the period to be predicted and the corresponding historical time series data of meteorology are input into the adapted photovoltaic power output prediction model to obtain the photovoltaic power output prediction result. The parameter adjustment of the basic prediction model also includes: Based on the ratio of the sample size at the second site to the sample size at the first site, select the parameter adjustment method; When the ratio is lower than a preset threshold, a freeze adjustment method is used to fix the bottom network parameters and update the top network parameters. When the ratio is not lower than a preset threshold, a full network adjustment method is adopted, and all network parameters are updated with the parameters of the basic prediction model as the initial values.

2. The photovoltaic power output prediction method based on VMD-feature transfer fusion in a data-scarce scenario as described in claim 1, characterized in that, The construction of a multi-dimensional feature set, based at least on the photovoltaic output time-series data of the first site and the corresponding meteorological time-series data, includes: Obtain the meteorological time-series data and time characteristic data of the first station; Multiple meteorological factors in the meteorological time-series data are combined and transformed to obtain derived meteorological characteristics; Lag characteristics are extracted from the photovoltaic power output time series data and the meteorological time series data to obtain lag time series features; The time feature data is converted into periodic encoding to obtain periodic encoded features; The multidimensional feature set is obtained by integrating the derived meteorological features, the lag time series features, and the periodic coding features.

3. The photovoltaic output prediction method based on VMD-feature transfer fusion in a data-scarce scenario as described in claim 1, characterized in that, The photovoltaic power output time series data is decomposed, including: Variational mode decomposition is used to decompose the photovoltaic power output time series data.

4. The photovoltaic output prediction method based on VMD-feature transfer fusion in a data-scarce scenario as described in claim 1, characterized in that, The step of determining the prediction result correction rule of the basic prediction model based on the first low-frequency component includes: The first low-frequency component is used as the daily cycle baseline trend of photovoltaic power output; Based on the daily cycle baseline trend, physical constraints are set for the photovoltaic power output prediction results to obtain the prediction result correction rules.

5. The photovoltaic output prediction method based on VMD-feature transfer fusion in a data-scarce scenario as described in claim 1, characterized in that, Establishing the mapping relationship between the multidimensional feature set and the second intermediate frequency component includes: The photovoltaic power output time series data and meteorological time series data of the first site are input into a long short-term memory network to obtain deep time series features; The deep temporal features are fused with the multidimensional feature set to obtain a fused feature set; Using the fused feature set as input and the second intermediate frequency component as the prediction target, the mapping relationship between the multidimensional feature set and the second intermediate frequency component is established.

6. The photovoltaic output prediction method based on VMD-feature transfer fusion in a data-scarce scenario as described in claim 1, characterized in that, The adaptive optimization is implemented using the particle swarm optimization algorithm.

7. The photovoltaic output prediction method based on VMD-feature transfer fusion in a data-scarce scenario as described in claim 1, characterized in that, The step of constructing a time-series prediction model corresponding to the third high-frequency component based on the meteorological time-series data and the third high-frequency component as input, according to the model parameters, includes: Configure the single-layer recurrent neural network structure of the time series prediction model according to the model parameters; Using the meteorological time-series data and the third high-frequency component as input, and combining the single-layer recurrent neural network structure, the time-series prediction model is constructed.

8. The photovoltaic output prediction method based on VMD-feature transfer fusion in a data-scarce scenario as described in claim 1, characterized in that, The step of adjusting the parameters of the basic prediction model to obtain an adapted photovoltaic output prediction model includes: Obtain the photovoltaic power output time-series data and the corresponding meteorological time-series data of the second site; Based on the same frequency band characteristics as the first site, the photovoltaic output time series data of the second site is decomposed into corresponding low-frequency components, mid-frequency components and high-frequency components; The model parameters of the mapping relationship are adjusted according to the mid-frequency component, and the model parameters of the time-series prediction model are adjusted according to the high-frequency component to obtain the adapted photovoltaic power output prediction model.

9. A photovoltaic power output prediction system based on VMD-feature transfer fusion in a data-scarce scenario, characterized in that, include: The feature construction module is used to construct a multi-dimensional feature set based at least on the photovoltaic power output time-series data of the first site and the corresponding meteorological time-series data; The frequency band decomposition module is used to decompose the photovoltaic power output time series data into a first low-frequency component, a second medium-frequency component, and a third high-frequency component according to the frequency band characteristics of the photovoltaic power output time series data. The correction rule determination module is used to determine the correction rule for the prediction result of the basic prediction model based on the first low-frequency component. The mapping relationship construction module is used to establish the mapping relationship between the multidimensional feature set and the second intermediate frequency component, and to determine the model parameters of the mapping relationship through adaptive optimization; The time series prediction modeling module is used to construct a time series prediction model corresponding to the third high frequency component based on the meteorological time series data and the third high frequency component as input and the model parameters. The basic model generation module is used to obtain the basic prediction model from the mapping relationship, the time series prediction model, and the prediction result correction rules; The model adaptation module is used to adjust the parameters of the basic prediction model based on the photovoltaic power output time series data of the second site and the corresponding meteorological time series data to obtain an adapted photovoltaic power output prediction model. The prediction execution module is used to input the historical time series data of photovoltaic power output of the second site before the period to be predicted and the corresponding historical time series data of meteorology into the adapted photovoltaic power output prediction model to obtain the photovoltaic power output prediction result. The parameter adjustment of the basic prediction model also includes: Based on the ratio of the sample size at the second site to the sample size at the first site, select the parameter adjustment method; When the ratio is lower than a preset threshold, a freeze adjustment method is used to fix the bottom network parameters and update the top network parameters. When the ratio is not lower than a preset threshold, a full network adjustment method is adopted, and all network parameters are updated with the parameters of the basic prediction model as the initial values.

Images