Photovoltaic power prediction method and system based on variational mode decomposition and time sequence modeling

Abstract

The invention relates to the technical field of photovoltaic power prediction, in particular to a photovoltaic power prediction method and system based on variational mode decomposition and time sequence modeling, and the method comprises the steps: obtaining a photovoltaic power sequence and meteorological observation data corresponding to the photovoltaic power sequence, and carrying out the missing value filling and normalization mapping processing; performing multi-scale decoupling on the processed power sequence through variational mode decomposition to obtain a plurality of mode functions; screening out an effective mode for the plurality of mode functions in combination with related contribution degree evaluation; constructing a modal weight function of time variation according to the correlation between the modals and the current power state, introducing a cross-modal shared convolution kernel structure into the time sequence convolution network, and performing dynamic weight fusion on effective modals of different time scales; explicitly embedding the modal weighting function into the conditional quantile prediction model; and predicting the photovoltaic power by using the trained conditional quantile prediction model. According to the method, fine description of the multi-time-scale dynamic characteristics and the uncertainty propagation law of the photovoltaic power is realized.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic power prediction technology, and in particular to a photovoltaic power prediction method and system based on variational mode decomposition and time series modeling. Background Technology

[0002] Solar photovoltaic (PV) power generation occupies a crucial position in my country's energy transition strategy due to its green and low-carbon nature, abundant resources, and mature technology. However, PV power generation is significantly affected by meteorological conditions such as solar irradiance, atmospheric cloud cover, temperature, and humidity, exhibiting strong randomness and non-stationarity, especially with dramatic and rapid fluctuations in the short term. If the uncertainties of PV power generation cannot be accurately characterized, it will adversely affect power system dispatch planning, reserve capacity allocation, and the ability to absorb new energy sources. Therefore, achieving high-precision and reliable short-term probabilistic prediction of PV power generation has significant engineering application value.

[0003] Photovoltaic power prediction technology can be broadly classified into two categories: deterministic prediction and probabilistic prediction. Deterministic prediction provides a single-point estimate and cannot reflect the distribution characteristics of the prediction error; probabilistic prediction, on the other hand, characterizes the prediction uncertainty through prediction intervals or probability distributions, and can provide a quantitative basis for scheduling strategies, thus becoming a research hotspot in recent years.

[0004] However, existing short-term probabilistic prediction methods for photovoltaic power have the following technical defects: they cannot effectively separate the multi-scale components of photovoltaic time series, resulting in enhanced coupling between high-frequency disturbances and low-frequency trends; the models lack agile response capabilities to rapid fluctuation scenarios, and the prediction stability is insufficient; the prediction interval coverage deviation is large, making it difficult to meet the power system's requirements for confidence reliability; and the model structure complexity is too high, making it difficult to achieve engineering deployment and real-time inference.

[0005] The information disclosed in this background section is intended only to enhance the understanding of the general background of the invention and should not be construed as an admission or in any way implying that the information constitutes prior art known to those skilled in the art. Summary of the Invention

[0006] This invention provides a photovoltaic power prediction method and system based on variational mode decomposition and time series modeling, thereby effectively solving the problems in the background technology.

[0007] To achieve the above objectives, the technical solution adopted by this invention is: a photovoltaic power prediction method based on variational mode decomposition and time series modeling, comprising the following steps:

[0008] Acquire photovoltaic power sequences and their corresponding meteorological observation data, and perform missing value imputation and normalization mapping;

[0009] The processed power sequence is decoupled at multiple scales by variational mode decomposition to obtain several mode functions; effective modes are selected from these mode functions by combining relevant contribution evaluation.

[0010] Based on the correlation between the mode and the current power state, a time-varying mode weight function is constructed. A cross-modal shared convolutional kernel structure is introduced into the temporal convolutional network to dynamically weight and fuse effective modes at different time scales.

[0011] The modal weighting function is explicitly embedded into the conditional quantile prediction model; the trained conditional quantile prediction model is used to predict photovoltaic power.

[0012] Furthermore, the missing value imputation and normalization mapping processing includes:

[0013] For moments on the timeline that are missing data, a linear interpolation method is used to recover them. The interpolation formula is as follows:

[0014] ;

[0015] In the formula, It is an interpolation node. The estimated value obtained by interpolation at this location and These represent the x and y coordinates of adjacent known observation points, respectively.

[0016] The input variables are linearly normalized to fall within the same numerical range. The normalization transformation is as follows:

[0017] ;

[0018] In the formula, These are the original observations. and These are the minimum and maximum values ​​of the variable in the sample, respectively. This is the result after normalization.

[0019] Furthermore, the power sequence after processing is decoupled at multiple scales through variational mode decomposition to obtain several mode functions, including:

[0020] Photovoltaic power output can be represented as a superposition of trend and disturbance components:

[0021] ;

[0022] In the formula, This indicates a low-frequency trend dominated by diurnal cycles and seasonal variations in irradiance. It reflects mid-frequency weather disturbances caused by changes in cloud cover, air mass movement, etc. This describes the high-frequency flicker caused by rapid movement of cloud edges and turbulence effects;

[0023] Variational mode decomposition is used to decompose the original sequence into multiple non-interfering narrowband eigenmode functions. The optimization problem is expressed as:

[0024] ;

[0025] In the formula, For the first One modal function, The center frequency of this mode. For time differential operators, The number of modes obtained from the decomposition.

[0026] Furthermore, the step of selecting effective modes from several modal functions by combining relevant contribution evaluation includes:

[0027] The formula for calculating the relevant contribution is:

[0028] ;

[0029] In the formula, For the first Individual modes and photovoltaic output Correlation indicators between them Represents the correlation coefficient operator;

[0030] when Less than the preset threshold At that time, if the modality is considered to have a low contribution to the prediction task, it is discarded, and only those that meet the requirements are retained. The modes constitute the effective mode set. .

[0031] Furthermore, the construction of the time-varying mode weight function based on the correlation between the mode and the current power state includes:

[0032] ;

[0033] In the formula, For the first The effective modes at time... Normalized weights, The number of valid modes to retain;

[0034] The modal weighting function is dynamically adjusted according to changes in weather and operating scenario. In scenarios where clouds obscure the view quickly, the weight of high-frequency modes is relatively increased, while under stable clear sky conditions, the weight of low-frequency trend modes dominates.

[0035] Furthermore, the introduction of a cross-modal shared convolutional kernel structure in the temporal convolutional network to dynamically weight and fuse effective modalities at different time scales includes:

[0036] The formula for calculating convolutional features is:

[0037] ;

[0038] In the formula, For at any time A unified temporal feature representation, The kernel length is 1. As the expansion factor, For the first Each convolutional kernel coefficient is shared across all modalities, i.e., for any modal index... All use the same .

[0039] Further, explicitly embedding the modal weight function into the conditional quantile prediction model includes:

[0040] The modality weight function Explicitly incorporating conditional quantile forecasting models allows confidence intervals to automatically adjust with weather changes; the quantile forecasting expression is expanded as follows:

[0041] ;

[0042] In the formula, For at any time Corresponding quantiles The predicted power value, For parameters A defined nonlinear mapping function, while relying on uniform features Effective modal weight set With target quantile .

[0043] Furthermore, in the process of using the trained conditional quantile prediction model to predict photovoltaic power, the model training includes:

[0044] The model training uses the consistency quantile loss function:

[0045] ;

[0046] In the formula, For actual observed power, These are the predicted values ​​for the corresponding quantiles. The target quantile;

[0047] Based on the prediction results of different quantiles, the upper and lower boundaries of the photovoltaic power prediction interval are constructed, and the lower quantile is set according to different confidence levels. With upper quantile The prediction interval is:

[0048] .

[0049] This invention also includes a photovoltaic power prediction system based on variational mode decomposition and time series modeling, using the method described above, wherein the system comprises:

[0050] The acquisition unit is used to acquire photovoltaic power sequences and their corresponding meteorological observation data, and to perform missing value imputation and normalization mapping.

[0051] The variational mode unit is used to perform multi-scale decoupling on the processed power sequence through variational mode decomposition to obtain several mode functions; and to select effective modes from the several mode functions by combining relevant contribution evaluation.

[0052] The temporally consistent unit is used to construct a time-varying modal weight function based on the correlation between the modality and the current power state. It introduces a cross-modal shared convolutional kernel structure into the temporal convolutional network to dynamically weight and fuse effective modes at different time scales.

[0053] The prediction unit is used to explicitly embed the modal weight function into the conditional quantile prediction model; and to predict photovoltaic power using the trained conditional quantile prediction model.

[0054] The present invention also includes a computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the method as described above.

[0055] The present invention also includes a storage medium having a computer program stored thereon, which, when executed by a processor, implements the method as described above.

[0056] The beneficial effects of this invention are as follows: By using variational mode decomposition and combining it with relevant contribution evaluation, multi-scale decoupling and effective mode selection of power sequences are achieved, theoretically reducing error coupling between disturbances in different frequency bands and solving the problem of excessive noise interference caused by mixing all time scales in traditional methods; by introducing a cross-modal shared convolutional kernel structure into the temporal convolutional network, effective modes at different time scales are uniformly represented by convolution, overcoming the problem of lacking a unified semantic space due to separate modeling of features at each scale, and improving the model's ability to characterize the co-evolution of high-frequency flicker and low-frequency trends; by constructing a modal weight function that is dynamically updated according to changes in meteorological conditions and explicitly embedding it into the conditional quantile regression model, this invention achieves adaptive shrinking or widening of the prediction interval width with the level of uncertainty, solving the problem that traditional static quantile models cannot simultaneously ensure the reliability and compactness of the interval in scenarios with rapid changes in weather conditions; combined with feature engineering design oriented towards physical mechanisms, the input features retain key information on irradiation-driven, component thermal effects, and atmospheric disturbances while forming a consistent representation space with multi-scale power modes, providing a solid data foundation for dynamic weight adjustment and consistent convolutional learning. Therefore, while ensuring the physical interpretability of the modeling process, this invention achieves a fine characterization of the dynamic characteristics and uncertainty propagation laws of photovoltaic power across multiple time scales. It has advantages such as high prediction accuracy, probability interval coverage close to the target confidence level, flexible response to changes in operating conditions, and applicability to grid-connected scheduling of high proportions of new energy sources. Attached Figure Description

[0057] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0058] Figure 1 This is a flowchart of the method in Example 1;

[0059] Figure 2 This is a schematic diagram of the system structure in Example 1;

[0060] Figure 3 This is a schematic diagram of the prediction model structure in Example 2;

[0061] Figure 4 This is a comparison chart of the predicted range and the measured photovoltaic power in Example 2;

[0062] Figure 5 This is a schematic diagram of the structure of the computer device of the present invention. Detailed Implementation

[0063] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0064] Example 1:

[0065] like Figure 1 As shown: A photovoltaic power prediction method based on variational mode decomposition and time series modeling includes the following steps:

[0066] Acquire photovoltaic power sequences and their corresponding meteorological observation data, and perform missing value imputation and normalization mapping;

[0067] The processed power sequence is decoupled at multiple scales by variational mode decomposition to obtain several mode functions; effective modes are selected from these mode functions by combining relevant contribution evaluation.

[0068] Based on the correlation between the mode and the current power state, a time-varying mode weight function is constructed. A cross-modal shared convolutional kernel structure is introduced into the temporal convolutional network to dynamically weight and fuse effective modes at different time scales.

[0069] The modal weighting function is explicitly embedded into the conditional quantile prediction model; the trained conditional quantile prediction model is used to predict photovoltaic power.

[0070] By employing variational mode decomposition and relevant contribution evaluation, multi-scale decoupling and effective mode selection of power sequences are achieved, theoretically reducing error coupling between perturbations in different frequency bands and solving the problem of excessive noise interference caused by mixing all time scales in traditional methods. By introducing a cross-modal shared convolutional kernel structure into the temporal convolutional network, effective modes at different time scales are uniformly represented by convolution, overcoming the problem of lacking a unified semantic space due to separate modeling of features at each scale, and improving the model's ability to characterize the co-evolution of high-frequency flicker and low-frequency trends. By constructing a modal weight function that is dynamically updated according to meteorological conditions and explicitly embedding it into the conditional quantile regression model, this invention achieves adaptive shrinking or widening of the prediction interval width according to the level of uncertainty, solving the problem that traditional static quantile models cannot simultaneously ensure the reliability and compactness of the interval in scenarios with rapid changes in weather conditions. Combined with feature engineering design oriented towards physical mechanisms, the input features retain key information on irradiation-driven, component thermal effects, and atmospheric perturbations while forming a consistent representation space with multi-scale power modes, providing a solid data foundation for dynamic weight adjustment and consistent convolutional learning.

[0071] Therefore, this embodiment, while ensuring the physical interpretability of the modeling process, achieves a fine characterization of the dynamic characteristics and uncertainty propagation laws of photovoltaic power across multiple time scales. It has advantages such as high prediction accuracy, probability interval coverage close to the target confidence level, flexible response to changes in operating conditions, and applicability to grid-connected scheduling of high proportions of new energy sources.

[0072] In this embodiment, missing value imputation and normalization mapping processing are performed, including:

[0073] For moments on the timeline that are missing data, a linear interpolation method is used to recover them. The interpolation formula is as follows:

[0074] ;

[0075] In the formula, It is an interpolation node. The estimated value obtained by interpolation at this location and These represent the x and y coordinates of adjacent known observation points, respectively.

[0076] The input variables are linearly normalized to fall within the same numerical range. The normalization transformation is as follows:

[0077] ;

[0078] In the formula, These are the original observations. and These are the minimum and maximum values ​​of the variable in the sample, respectively. This is the result after normalization.

[0079] Among them, the processed power sequence is decoupled at multiple scales through variational mode decomposition to obtain several mode functions, including:

[0080] Photovoltaic power output can be represented as a superposition of trend and disturbance components:

[0081] ;

[0082] In the formula, This indicates a low-frequency trend dominated by diurnal cycles and seasonal variations in irradiance. It reflects mid-frequency weather disturbances caused by changes in cloud cover, air mass movement, etc. This describes the high-frequency flicker caused by rapid movement of cloud edges and turbulence effects;

[0083] Variational mode decomposition is used to decompose the original sequence into multiple non-interfering narrowband eigenmode functions. The optimization problem is expressed as:

[0084] ;

[0085] In the formula, For the first One modal function, The center frequency of this mode. For time differential operators, The number of modes obtained from the decomposition.

[0086] : Effective modes are selected from several modal functions by combining relevant contribution evaluations, including:

[0087] The formula for calculating the relevant contribution is:

[0088] ;

[0089] In the formula, For the first Individual modes and photovoltaic output Correlation indicators between them Represents the correlation coefficient operator;

[0090] when Less than the preset threshold At that time, if the modality is considered to have a low contribution to the prediction task, it is discarded, and only those that meet the requirements are retained. The modes constitute the effective mode set. .

[0091] As a preferred embodiment of the above, constructing a time-varying mode weighting function based on the correlation between the mode and the current power state includes:

[0092] ;

[0093] In the formula, For the first The effective modes at time... Normalized weights, The number of valid modes to retain;

[0094] The modal weighting function is dynamically adjusted according to changes in weather and operating scenario. In scenarios where clouds obscure the view quickly, the weight of high-frequency modes is relatively increased, while under stable clear sky conditions, the weight of low-frequency trend modes dominates.

[0095] In this embodiment, a cross-modal shared convolutional kernel structure is introduced into the temporal convolutional network to dynamically weight and fuse effective modalities at different time scales, including:

[0096] The formula for calculating convolutional features is:

[0097] ;

[0098] In the formula, For at any time A unified temporal feature representation, The kernel length is 1. As the expansion factor, For the first Each convolutional kernel coefficient is shared across all modalities, i.e., for any modal index... All use the same .

[0099] The explicit embedding of modal weight functions into the conditional quantile prediction model includes:

[0100] modal weight function Explicitly incorporating conditional quantile forecasting models allows confidence intervals to automatically adjust with weather changes; the quantile forecasting expression is expanded as follows:

[0101] ;

[0102] In the formula, For at any time Corresponding quantiles The predicted power value, For parameters A defined nonlinear mapping function, while relying on uniform features Effective modal weight set With target quantile .

[0103] In predicting photovoltaic power using the trained conditional quantile prediction model, model training includes:

[0104] The model training uses the consistency quantile loss function:

[0105] ;

[0106] In the formula, For actual observed power, These are the predicted values ​​for the corresponding quantiles. The target quantile;

[0107] Based on the prediction results of different quantiles, the upper and lower boundaries of the photovoltaic power prediction interval are constructed, and the lower quantile is set according to different confidence levels. With upper quantile The prediction interval is:

[0108] .

[0109] like Figure 2As shown, this embodiment also includes a photovoltaic power prediction system based on variational mode decomposition and time series modeling. Using the method described above, the system includes:

[0110] The acquisition unit is used to acquire photovoltaic power sequences and their corresponding meteorological observation data, and to perform missing value imputation and normalization mapping.

[0111] Variational mode units are used to decouple the processed power sequence at multiple scales through variational mode decomposition to obtain several mode functions; and effective modes are selected from these mode functions by combining relevant contribution evaluation.

[0112] The temporally consistent unit is used to construct a time-varying modal weight function based on the correlation between the modality and the current power state. It introduces a cross-modal shared convolutional kernel structure into the temporal convolutional network to dynamically weight and fuse effective modes at different time scales.

[0113] The prediction unit is used to explicitly embed the modal weighting function into the conditional quantile prediction model; the trained conditional quantile prediction model is used to predict photovoltaic power.

[0114] Example 2:

[0115] This embodiment provides a photovoltaic power prediction method based on variational mode decomposition and time series modeling. The model structure is as follows: Figure 3 As shown, it includes the following steps:

[0116] Step 1: Raw data processing.

[0117] To ensure the computability and numerical stability of the subsequent modeling process, this embodiment first performs missing value imputation and normalization mapping on the photovoltaic power sequence and its corresponding meteorological observations. For moments with missing data on the time axis, a linear interpolation method is used for recovery, and the interpolation formula can be written as:

[0118] (1)

[0119] In the formula, It is an interpolation node. The estimated value obtained by interpolation at this location and These represent the x and y coordinates of adjacent known observation points, respectively. After interpolation, each input variable is linearly normalized to ensure it falls within the same numerical range. The normalization transformation is as follows:

[0120] (2)

[0121] In the formula, These are the original observations. and These are the minimum and maximum values ​​of the variable in the sample, respectively. This is the result after normalization. Through this preprocessing, the original signal is mapped into a bounded, continuous time series with consistent dimensions, providing a convergent and comparable basis for subsequent variational optimization and convolution operations.

[0122] Step 2: Multi-scale signal decoupling based on variational mode decomposition.

[0123] Photovoltaic power output can be abstractly represented as the superposition of trend components and disturbance components, that is:

[0124] (3)

[0125] In the formula, This indicates a low-frequency trend dominated by diurnal cycles and seasonal variations in irradiance. It reflects mid-frequency weather disturbances caused by changes in cloud cover, air mass movement, etc. This describes the high-frequency flicker caused by rapid cloud edge movement and turbulence effects. Without decomposition, directly... If a holistic model is performed, the prediction error term will be... It contains multiple sources, and its variance is not only composed of the variances of each component error, but also affected by the covariance between different error components, which can be written as:

[0126] (4)

[0127] When a significant covariance term exists, the lower bound of the error variance increases, which theoretically limits the optimal accuracy achievable by the model. To reduce the coupling of errors, this invention employs variational mode decomposition to decompose the original sequence into multiple non-interfering narrowband eigenmode functions. Its optimization problem can be expressed as:

[0128] (5)

[0129] In the formula, For the first One modal function, The center frequency of this mode. For time differential operators, The number of modes obtained by decomposition is given. Through variational constraints and the narrowband assumption, variational mode decomposition projects the original mixed signal into a set of relatively concentrated frequency subspaces. The frequency domain overlap between different modes is significantly suppressed, so that the error term can be approximately written as the sum of the errors of each mode. The covariance term is significantly reduced, and the upper bound of the error variance is thus lowered.

[0130] After obtaining each mode, this invention filters effective modes based on the correlation contribution between the mode and the target power sequence. The correlation contribution calculation formula is as follows:

[0131] (6)

[0132] In the formula, For the first Individual modes and photovoltaic output Correlation indicators between them This represents the correlation coefficient operator. When... Less than the preset threshold At that time, if the modality is considered to contribute little to the prediction task, it is discarded, and only those that meet the requirements are retained. The modes constitute the effective mode set. This step projects multi-scale dynamics into a subspace highly relevant to the prediction task, theoretically reducing error coupling between different noise sources.

[0133] Step 3: Consistent temporal modeling based on dynamic weighting and shared convolutional kernels.

[0134] After obtaining an effective set of modes, this invention proposes a consistent temporal modeling mechanism based on dynamic mode weights and shared convolutional kernels, representing dynamic changes at different time scales through convolution within a unified semantic space. First, a time-varying mode weight function is constructed based on the correlation between the mode and the current power state, and its expression is:

[0135] (7)

[0136] In the formula, For the first The effective modes at time... Normalized weights, This represents the number of valid modes to retain. This weight is dynamically adjusted according to weather and operating scenarios. In scenarios such as rapid cloud cover, the weight of high-frequency modes is relatively increased, while under stable clear sky conditions, the weight of low-frequency trend modes dominates, thus reflecting the physical mechanism of photovoltaic output changes.

[0137] Based on this, this invention utilizes dilated convolutional structures to perform unified temporal modeling of various modalities, innovatively introducing cross-modal shared convolutional kernels, and weighted fusion using dynamic weights. The calculation formula for convolutional features is as follows:

[0138] (8)

[0139] In the formula, For at any time A unified temporal feature representation, The kernel length is 1. As the expansion factor, For the first Each convolutional kernel coefficient is shared across all modalities, i.e., for any modal index... All use the same Through this structural design, modalities at different time scales are mapped to a unified temporal semantic space under the same set of convolutional kernels and dynamic weight modulation. This preserves multi-scale information and ensures cross-scale dynamic consistency, which is beneficial to improving the model's prediction stability and generalization ability in scenarios with drastic power fluctuations.

[0140] Step 4: Conditional quantile regression probability prediction based on modal weights:

[0141] In obtaining a unified temporal feature representation Subsequently, this invention enhances the interval adaptive capability under different meteorological conditions based on a prediction mechanism involving both modal weights and confidence quantiles. This is considering the dynamic weights. This invention can reflect the contribution of disturbances at different time scales to photovoltaic power in the current scenario. Explicitly incorporating conditional quantile forecasting models allows confidence intervals to automatically adjust for weather changes. The quantile forecasting expression is expanded as follows:

[0142] (9)

[0143] In the formula, For at any time Corresponding quantiles The predicted power value, For parameters A defined nonlinear mapping function, while relying on uniform features Effective modal weight set With target quantile Unlike existing quantile regression methods that rely on only a single feature input, this invention introduces... As additional prior information, this enables the model to have dynamic adaptive capabilities in non-stationary scenarios such as rapid changes in weather conditions.

[0144] The model training still uses the consistency quantile loss function:

[0145] (10)

[0146] In the formula, For actual observed power, These are the predicted values ​​for the corresponding quantiles. The target quantile is used. Based on the prediction results of different quantiles, the upper and lower boundaries of the photovoltaic power prediction interval can be constructed. For example, at a confidence level of 90%, the lower quantile can be selected. With upper quantile The prediction interval is written as:

[0147] (11)

[0148] Based on the dynamic modal weight update mechanism, the high-frequency modal weights increase under rapid shadowing perturbations, causing the prediction interval to widen to reflect the increased risk; under stable clear sky conditions, the low-frequency modal weights dominate, and the prediction interval converges and narrows, reflecting high confidence accuracy. Therefore, this invention achieves a responsive control mechanism for uncertainty in probabilistic prediction that adapts to changes in the physical environment, significantly outperforming existing static interval modeling methods with fixed structures.

[0149] To verify the applicability of the method of this invention in actual photovoltaic scenarios, the invention was applied to measured data from a grid-connected photovoltaic power station. The power station has an installed capacity of 40 MW and a sampling resolution of 15 minutes. Six consecutive months of photovoltaic operation data were selected as the research sample. Input variables included historical photovoltaic power and various meteorological characteristics coupled with short-term power generation dynamics, such as global horizontal irradiance, diffuse irradiance, ambient temperature, relative humidity, 10 m and 100 m wind speeds and directions, cloud cover, and atmospheric pressure. These variables were strictly time-synchronized and then input into the model.

[0150] Before model training, this invention employs physical mechanism-oriented feature engineering design for multi-source input data to ensure that the input information is both comprehensive and computable. All power and meteorological observation sequences undergo missing data restoration and noise suppression, and scale bias caused by different physical dimensions is reduced through normalization mapping, thereby improving the model's training stability and gradient propagation efficiency. In feature construction, this invention selects key driving factors based on the short-term dynamic laws of photovoltaic power generation, considering irradiance energy input, module thermal effects, and atmospheric disturbance mechanisms, and fuses these factors with the decoupled effective power modes from multiple scales to form a unified input. The fused features not only possess clear physical interpretability but also retain the ability to express differentiated high-frequency flicker and low-frequency trends, providing a reliable prior foundation for the dynamic mode weight adjustment and consistent convolutional learning proposed in this invention, and helping to enhance the model's sensitivity to prediction uncertainties under different meteorological conditions.

[0151] The data is strictly divided into training and testing sets according to chronological order. The first 80% of continuous samples are used for model parameter learning, and the last 20% are used for generalization performance evaluation. This avoids overestimation of performance due to future information leakage and ensures that the experimental process meets the engineering application requirements of short-term power system forecasting. Subsequently, this embodiment performs data preprocessing, multi-scale modal decomposition and screening, dynamic consistency convolution modeling, and conditional quantile regression with modal weights in sequence according to steps 1 to 4 to achieve probabilistic interval prediction of photovoltaic power.

[0152] In the test set, the method of this invention was compared and evaluated with several mainstream probabilistic prediction methods. Table 1 shows the performance of different models in terms of CRPS, PICP90, and PINAW90 at the 90% prediction interval. This invention achieves superior performance in Continuous Rank Probability Score (CRPS), indicating that it can more accurately characterize the prediction distribution; the PICP index is close to the target confidence level, demonstrating good reliability of the probability interval.

[0153] at the same time, Figure 4 The comparison between the predicted range and the measured power of the method of this invention is presented. It can be seen that the overall coverage of the predicted range is reasonable and consistent with the actual power change trend. When the power level changes rapidly, the predicted range can be adjusted accordingly, demonstrating its adaptive characterization ability to changes in uncertainty; when the power change is relatively gradual, the range width tends to converge, indicating that the model can maintain the tightness of the prediction results while ensuring reliability. These results fully demonstrate that the present invention has advantages in stability and interpretability in engineering applications.

[0154] Table 1. Evaluation results of the 90% prediction interval probability performance of the method of the present invention after interval calibration.

[0155]

[0156] Please see Figure 5 The diagram shows a structural schematic of a computer device provided in an embodiment of this application. An embodiment of this application provides a computer device 400, including a processor 410 and a memory 420. The memory 420 stores a computer program executable by the processor 410. When the computer program is executed by the processor 410, it performs the method described above.

[0157] This application embodiment also provides a storage medium 430, on which a computer program is stored, and the computer program is executed by a processor 410 to perform the above method.

[0158] The storage medium 430 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0159] In the description of this invention, the terms "first" and "second" 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. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. "A plurality of" means two or more, unless otherwise explicitly specified.

[0160] In this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," "linking," and "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0161] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Furthermore, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0162] Any process or method description in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing a particular logical function or process, and the scope of the preferred embodiments of the invention includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as will be understood by those skilled in the art to which embodiments of the invention pertain.

[0163] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a ordered list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable media include: an electrical connection having one or more wires (electronic device), a portable computer disk drive (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Alternatively, the computer-readable medium may be paper or other suitable media on which the program can be printed, since the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in a computer memory.

[0164] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0165] Those skilled in the art will understand that all or part of the steps of the methods in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

[0166] The storage medium mentioned above can be a read-only memory, a disk, or an optical disk, etc. Although embodiments of the present invention have been shown and described above, it is to be understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of the present invention.

Claims

1. A photovoltaic power prediction method based on variational mode decomposition and time series modeling, characterized in that, Includes the following steps: Acquire photovoltaic power sequences and their corresponding meteorological observation data, and perform missing value imputation and normalization mapping; The processed power sequence is decoupled at multiple scales by variational mode decomposition to obtain several mode functions; effective modes are selected from these mode functions by combining relevant contribution evaluation. Based on the correlation between the mode and the current power state, a time-varying mode weight function is constructed. A cross-modal shared convolutional kernel structure is introduced into the temporal convolutional network to dynamically weight and fuse effective modes at different time scales. The modal weighting function is explicitly embedded into the conditional quantile prediction model; the trained conditional quantile prediction model is used to predict photovoltaic power.

2. The photovoltaic power prediction method based on variational mode decomposition and time series modeling according to claim 1, characterized in that, The missing value imputation and normalization mapping process includes: For moments on the timeline that are missing data, a linear interpolation method is used to recover them. The interpolation formula is as follows: ； In the formula, It is an interpolation node. The estimated value obtained by interpolation at this location and These represent the x and y coordinates of adjacent known observation points, respectively. The input variables are linearly normalized to fall within the same numerical range. The normalization transformation is as follows: ； In the formula, These are the original observations. and These are the minimum and maximum values ​​of the variable in the sample, respectively. This is the result after normalization.

3. The photovoltaic power prediction method based on variational mode decomposition and time series modeling according to claim 1, characterized in that, The power sequence after processing is decoupled at multiple scales through variational mode decomposition to obtain several mode functions, including: Photovoltaic power output can be represented as a superposition of trend and disturbance components: ； In the formula, This indicates a low-frequency trend dominated by diurnal cycles and seasonal variations in irradiance. It reflects mid-frequency weather disturbances caused by changes in cloud cover, air mass movement, etc. This describes the high-frequency flicker caused by rapid movement of cloud edges and turbulence effects; Variational mode decomposition is used to decompose the original sequence into multiple non-interfering narrowband eigenmode functions. The optimization problem is expressed as: ； In the formula, For the first One modal function, The center frequency of this mode. For time differential operators, The number of modes obtained from the decomposition.

4. The photovoltaic power prediction method based on variational mode decomposition and time series modeling according to claim 3, characterized in that, The process of selecting effective modes from several modal functions by combining relevant contribution evaluations includes: The formula for calculating the relevant contribution is: ； In the formula, For the first Individual modes and photovoltaic output Correlation indicators between them Represents the correlation coefficient operator; when Less than the preset threshold At that time, if the modality is considered to have a low contribution to the prediction task, it is discarded, and only those that meet the requirements are retained. The modes constitute the effective mode set. .

5. The photovoltaic power prediction method based on variational mode decomposition and time series modeling according to claim 4, characterized in that, The construction of the time-varying mode weight function based on the correlation between the mode and the current power state includes: ； In the formula, For the first The effective modes at time... Normalized weights, The number of valid modes to retain; The modal weighting function is dynamically adjusted according to changes in weather and operating scenario. In scenarios where clouds obscure the view quickly, the weight of high-frequency modes is relatively increased, while under stable clear sky conditions, the weight of low-frequency trend modes dominates.

6. The photovoltaic power prediction method based on variational mode decomposition and time series modeling according to claim 5, characterized in that, The method of introducing a cross-modal shared convolutional kernel structure into a temporal convolutional network to dynamically weight and fuse effective modalities at different time scales includes: The formula for calculating convolutional features is: ； In the formula, For at any time A unified temporal feature representation, The kernel length is 1. As the expansion factor, For the first Each convolutional kernel coefficient is shared across all modalities, i.e., for any modal index... All use the same .

7. The photovoltaic power prediction method based on variational mode decomposition and time series modeling according to claim 1, characterized in that, The explicit embedding of the modality weight function into the conditional quantile prediction model includes: The modality weight function Explicitly incorporating conditional quantile forecasting models allows confidence intervals to automatically adjust with weather changes; the quantile forecasting expression is expanded as follows: ； In the formula, For at any time Corresponding quantiles The predicted power value, For parameters A defined nonlinear mapping function, while relying on uniform features Effective modal weight set With target quantile .

8. The photovoltaic power prediction method based on variational mode decomposition and time series modeling according to claim 1, characterized in that, In the process of using the trained conditional quantile prediction model to predict photovoltaic power, model training includes: The model training uses the consistency quantile loss function: ； In the formula, For actual observed power, These are the predicted values ​​for the corresponding quantiles. The target quantile; Based on the prediction results of different quantiles, the upper and lower boundaries of the photovoltaic power prediction interval are constructed, and the lower quantile is set according to different confidence levels. With upper quantile The prediction interval is: 。 9. A photovoltaic power prediction system based on variational mode decomposition and time series modeling, characterized in that, Using the method of any one of claims 1 to 8, the system comprises: The acquisition unit is used to acquire photovoltaic power sequences and their corresponding meteorological observation data, and to perform missing value imputation and normalization mapping. The variational mode unit is used to perform multi-scale decoupling on the processed power sequence through variational mode decomposition to obtain several mode functions; and to select effective modes from the several mode functions by combining relevant contribution evaluation. The temporally consistent unit is used to construct a time-varying modal weight function based on the correlation between the modality and the current power state. It introduces a cross-modal shared convolutional kernel structure into the temporal convolutional network to dynamically weight and fuse effective modes at different time scales. The prediction unit is used to explicitly embed the modal weight function into the conditional quantile prediction model; and to predict photovoltaic power using the trained conditional quantile prediction model.

10. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method as described in any one of claims 1-8.

11. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1-8.

Images