Deep learning-based weather-aware heterogeneous dual-flow decoupling prediction method and device

Abstract

This disclosure provides an exemplary embodiment of a deep learning-based weather perception heterogeneous dual-flow decoupling prediction method, apparatus, device, medium, and program product. The method includes: acquiring raw data and calculating solar position characteristics and theoretical clear-sky irradiance to construct a meteorological matrix; generating a reliability score and dynamic fusion weights based on the deviation between weather forecast values ​​and measured values; decomposing historical photovoltaic power sequences into trend components and fluctuation components and embedding them into blocks respectively; performing heterogeneous dual-flow parallel modeling based on first and second neural networks to obtain trend branches and fluctuation branches, and fusing the block embedding results into the two branches; determining the current weather state, generating dynamic weights for the two branches, obtaining predicted values ​​for each branch, weighted and fused, and outputting a future photovoltaic power prediction sequence. This disclosure improves the accuracy of photovoltaic power prediction under different weather conditions through weather heterogeneous dual-flow decoupling prediction.

Description

Technical Field

[0001] This disclosure relates to the field of photovoltaic power prediction technology, and in particular to a method, apparatus, equipment, medium and program product for weather perception heterogeneous dual-stream decoupling prediction based on deep learning. Background Technology

[0002] Currently, photovoltaic power prediction based on weather forecasts relies too heavily on these forecasts, and low-quality forecasts can severely interfere with the accuracy of photovoltaic power prediction. Deep learning-based predictions rely directly on raw observation data and historical data, lacking physical knowledge such as solar motion, resulting in low data utilization efficiency. Furthermore, using a single network to simultaneously fit low-frequency trend components and high-frequency fluctuation components of photovoltaic power leads to mutual interference in the fittings, resulting in different prediction performance under different weather scenarios. Meanwhile, predictions based on sequence decomposition mostly use fixed filters, without combining them with weather conditions or adjusting weights according to weather conditions, resulting in limited adaptability of the model under different weather scenarios. Summary of the Invention

[0003] To address the aforementioned technical problems, or at least partially address them, this disclosure provides a method, apparatus, device, medium, and program product for weather perception heterogeneous dual-flow decoupling prediction based on deep learning.

[0004] This exemplary embodiment provides a deep learning-based weather sensing heterogeneous dual-flow decoupling prediction method, the method comprising: The original data for the prediction time domain is determined, including historical photovoltaic power sequences, historical meteorological observation sequences, future meteorological forecast sequences, power plant latitude and longitude, and timestamps; Based on the original data, physical features are calculated to determine the solar position features and theoretical clear-sky irradiance. The historical meteorological observation sequence, the solar position features, the theoretical clear-sky irradiance, and the time period code are then spliced ​​together to construct a meteorological matrix. Based on the deviation between meteorological forecasts and actual meteorological values, a reliability score and dynamic fusion weights are generated. Based on the historical photovoltaic power sequence, the trend component and fluctuation component are determined. Through power dual-component block embedding, a power block embedding sequence is obtained. The meteorological matrix is ​​then block-embedded to obtain a meteorological block embedding sequence. Heterogeneous dual-stream parallel modeling is performed on the power block embedding sequence based on the first and second neural networks to obtain trend branches and fluctuation branches. Based on the dynamic fusion weights and cross attention, the meteorological block embedding sequence is fused into the two branches. Based on the physical characteristics, the current weather condition is calculated and determined. The dynamic weights of the two branches are adaptively generated, the predicted values ​​of each branch are obtained and weighted and fused, and the future photovoltaic power prediction sequence is output.

[0005] In some exemplary embodiments, the historical meteorological observation sequence includes at least total horizontal radiation, ambient temperature, wind speed, and relative humidity; The timestamps include Julian Day and True Solar Time; The meteorological matrix includes a historical enhanced meteorological matrix and a future meteorological matrix; Based on the raw data, physical feature calculations are performed to determine solar position characteristics and theoretical clear-sky irradiance. A meteorological matrix is ​​constructed by concatenating the historical meteorological observation sequence, the solar position characteristics, the theoretical clear-sky irradiance, and the time period encoding. This matrix includes: Physical feature calculations are performed based on the raw data. Based on the latitude and longitude of the power station and the timestamp, the solar position features are determined, including the solar altitude angle and the solar azimuth angle. Based on the aforementioned solar position characteristics, the theoretical clear-sky irradiance is determined; Based on the theoretical clear-sky irradiance, the clear-sky index is determined; The historical enhanced meteorological matrix is ​​constructed by splicing together the historical meteorological observation sequence, the solar position characteristics, the theoretical clear sky irradiance, the clear sky index, and the time period code. Based on the meteorological forecasting system, the future meteorological forecast sequence within the forecast time domain is determined, and based on the future meteorological forecast sequence and the theoretical clear sky irradiance, the predicted clear sky index is determined. The future weather matrix is ​​constructed by splicing together the solar position characteristics, the theoretical clear sky irradiance, the predicted clear sky index, and the time period code.

[0006] In some exemplary embodiments, a confidence score and dynamic fusion weight are generated based on the deviation between the weather forecast value and the observed weather value, including: Based on the theoretical clear-sky irradiance, the deviation between the weather forecast value and the actual weather measurement value at each time point within the historical window is determined; Based on the aforementioned deviation, the overall error is determined by exponential decay weighting. Based on the comprehensive error, the reliability score of the current weather forecast is determined through exponential mapping transformation; The confidence score, the comprehensive error, and the predicted clear sky index are concatenated and input into a gating network to generate dynamic fusion weights.

[0007] In some exemplary embodiments, the historical photovoltaic power sequence is decomposed to determine trend components and fluctuation components. A power block embedding sequence is obtained through power dual-component block embedding. The meteorological matrix is ​​then block-embedded to obtain a meteorological block embedding sequence, including: Based on the historical photovoltaic power sequence, the low-frequency trend component and high-frequency fluctuation component are determined by exponential moving average decomposition. Based on the historical enhanced meteorological matrix and the future meteorological matrix, the corresponding trend component and fluctuation component are determined through the exponential moving average decomposition. The corresponding trend component and fluctuation component are respectively divided into blocks along the time axis to obtain a block sequence; Based on the block sequence, the block is embedded by mapping to a latent space of a preset dimension through a linear layer, and a positional encoding corresponding to the original time position of the block sequence is applied to obtain a meteorological block embedding sequence. The low-frequency trend component and the high-frequency fluctuation component are respectively divided into blocks to obtain their respective power block sequences. The corresponding power block embedding sequence is obtained by power dual-component block embedding.

[0008] In some exemplary embodiments, heterogeneous dual-stream parallel modeling is performed on the power block embedding sequence based on the first and second neural networks to obtain a trend branch and a fluctuation branch. Based on the dynamic fusion weights and cross attention, the meteorological block embedding sequence is fused into the two branches, including: Based on the first neural network, the power block embedding sequence corresponding to the low-frequency trend component is processed to obtain the trend branch; Based on the second neural network, the power block embedding sequence corresponding to the high-frequency fluctuation component is processed to obtain the fluctuation branch; Based on the fluctuation branch, a query is performed, using the historical meteorological block embedding sequence and the future meteorological block embedding sequence as keys respectively. Two independent multi-head cross attention is executed, and the future meteorological block embedding sequence is adaptively controlled based on the dynamic fusion weight. The fused fluctuation branch is obtained through the first neural network. The multi-head cross attention is replaced by a lightweight linear weighted fusion to obtain the fused trend branch.

[0009] In some exemplary embodiments, the current weather state is calculated and determined based on the physical characteristics, the dynamic weights of the two branches are adaptively generated, the predicted values ​​of each branch are obtained and weighted and fused, and the future photovoltaic power prediction sequence is output, including: The current weather state is determined based on the physical characteristics. The weather state feature vector is determined based on the clear sky index, the average cloud cover forecast in the prediction time domain, and the confidence score. The weather state includes sunny, cloudy, and rainy days. The weather state feature vector is input into a linear weight generation network, and normalized by a normalized exponential function to obtain the dynamic weights of the fused trend branch and the dynamic weights of the fused fluctuation branch. A linear decoder is applied to the fused trend branch and the fused fluctuation branch respectively to obtain the predicted values ​​of the trend branch and the fluctuation branch. Based on the dynamic weights of the fused trend branch and the fused fluctuation branch, a weighted fusion is performed to obtain the future photovoltaic power prediction sequence.

[0010] This exemplary embodiment also provides a weather sensing heterogeneous dual-flow decoupling prediction device based on deep learning, the device comprising: The determination module is configured to determine the raw data in the prediction time domain, which includes historical photovoltaic power sequences, historical meteorological observation sequences, future meteorological forecast sequences, power plant latitude and longitude, and timestamps; The construction module is configured to perform physical feature calculations based on the original data, determine the solar position features and theoretical clear sky irradiance, and construct a meteorological matrix by splicing together the historical meteorological observation sequence, the solar position features, the theoretical clear sky irradiance, and the time period code. The generation module is configured to generate a credibility score and dynamic fusion weights based on the deviation between the weather forecast and the actual weather value. The embedding module is configured to decompose the historical photovoltaic power sequence to determine the trend component and fluctuation component, obtain the power block embedding sequence through power dual-component block embedding, and perform block embedding on the meteorological matrix to obtain the meteorological block embedding sequence. The fusion module is configured to perform heterogeneous dual-stream parallel modeling on the power block embedding sequence based on the first and second neural networks to obtain trend branches and fluctuation branches, and to fuse the meteorological block embedding sequence into the two branches based on the dynamic fusion weights and cross attention. The output module is configured to calculate and determine the current weather state based on the physical characteristics, adaptively generate the dynamic weights of the two branches, obtain the predicted values ​​of each branch and weighted fuse them, and output the future photovoltaic power prediction sequence.

[0011] An exemplary embodiment of this disclosure also provides an electronic device, the electronic device comprising: a processor; a memory for storing executable instructions of the processor; the processor being configured to read the executable instructions from the memory and execute the instructions to implement the deep learning-based weather perception heterogeneous dual-flow decoupling prediction method provided in the exemplary embodiment of this disclosure.

[0012] This disclosure also provides a computer-readable storage medium storing a computer program for performing a deep learning-based weather sensing heterogeneous dual-flow decoupling prediction method as provided in this disclosure.

[0013] The technical solution provided by the exemplary embodiments of this disclosure has the following advantages compared with the prior art: The heterogeneous dual-flow decoupling prediction method based on deep learning provided in this exemplary embodiment of the present disclosure constructs a meteorological matrix by acquiring raw data and calculating solar position characteristics and theoretical clear-sky irradiance; generating a confidence score and dynamic fusion weights based on the deviation between meteorological forecasts and actual meteorological values; decomposing historical photovoltaic power sequences into trend components and fluctuation components and embedding them into blocks respectively; performing parallel modeling of heterogeneous dual flows based on first and second neural networks to obtain trend branches and fluctuation branches, and fusing the block embedding results into the two branches; determining the current weather state, generating dynamic weights for the two branches, obtaining the predicted values ​​of each branch, weighted and fused, and outputting a future photovoltaic power prediction sequence. This disclosure improves the accuracy of photovoltaic power prediction under different weather conditions through heterogeneous dual-flow decoupling prediction. Attached Figure Description

[0014] The above and other features, advantages, and aspects of the embodiments of this disclosure will become more apparent from the accompanying drawings and the following detailed description. Throughout the drawings, the same or similar reference numerals denote the same or similar elements. It should be understood that the drawings are schematic, and the originals and elements are not necessarily drawn to scale.

[0015] Figure 1 A flowchart illustrating a deep learning-based weather perception heterogeneous dual-flow decoupling prediction method provided for an exemplary embodiment of this disclosure; Figure 2 A schematic diagram of the technical architecture for weather perception-based heterogeneous dual-flow structure prediction provided as an exemplary embodiment of this disclosure; Figure 3 A schematic diagram of the structure of a deep learning-based weather sensing heterogeneous dual-flow decoupling prediction device provided as an exemplary embodiment of this disclosure; Figure 4 A schematic diagram of the structure of an electronic device provided for an exemplary embodiment of this disclosure. Detailed Implementation

[0016] Embodiments of this disclosure will now be described in more detail with reference to the accompanying drawings. While some embodiments of this disclosure are shown in the drawings, it should be understood that this disclosure can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of this disclosure. It should be understood that the accompanying drawings and embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of protection of this disclosure.

[0017] It should be understood that the steps described in the method embodiments of this disclosure may be performed in different orders and / or in parallel. Furthermore, the method embodiments may include additional steps and / or omit the steps shown. The scope of this disclosure is not limited in this respect.

[0018] The term "comprising" and its variations as used herein are open-ended inclusions, meaning "including but not limited to". The term "based on" means "at least partially based on". The term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one additional embodiment"; the term "some embodiments" means "at least some embodiments". Definitions of other terms will be given in the description below.

[0019] It should be noted that the concepts of "first" and "second" mentioned in this disclosure are used only to distinguish different devices, modules or units, and are not used to limit the order of functions performed by these devices, modules or units or their interdependencies.

[0020] It should be noted that the terms "a" and "a plurality of" used in this disclosure are illustrative rather than restrictive, and those skilled in the art should understand that, unless otherwise expressly indicated in the context, they should be understood as "one or more".

[0021] The names of messages or information exchanged between multiple devices in the embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of such messages or information.

[0022] To address the aforementioned issues, exemplary embodiments of this disclosure provide a deep learning-based weather perception heterogeneous dual-flow decoupling prediction method, which will be described below with reference to specific embodiments.

[0023] Figure 1 This is a flowchart illustrating a deep learning-based weather sensing heterogeneous dual-flow decoupling prediction method for exemplary embodiments of the present disclosure. This method can be executed by a deep learning-based weather sensing heterogeneous dual-flow decoupling prediction device, which can be implemented in software and / or hardware and is generally integrated into an electronic device.

[0024] like Figure 1 , Figure 2As shown, the deep learning-based weather perception heterogeneous dual-flow decoupling prediction method includes the following steps: Step S101: Determine the raw data for the prediction time domain. The raw data includes historical photovoltaic power sequences, historical meteorological observation sequences, future meteorological forecast sequences, power station latitude and longitude, and timestamps. Step S102: Based on the original data, perform physical feature calculations to determine the solar position features and theoretical clear-sky irradiance. Then, based on the historical meteorological observation sequence, the solar position features, the theoretical clear-sky irradiance, and the time period code, splice them together to construct a meteorological matrix.

[0025] In some exemplary embodiments, the historical meteorological observation sequence includes at least total horizontal radiation (GHI), ambient temperature, wind speed, and relative humidity; The timestamps include Julian Day and True Solar Time; The meteorological matrix includes a historical enhanced meteorological matrix and a future meteorological matrix; Based on the raw data, physical feature calculations are performed to determine solar position characteristics and theoretical clear-sky irradiance. A meteorological matrix is ​​constructed by concatenating the historical meteorological observation sequence, the solar position characteristics, the theoretical clear-sky irradiance, and the time period encoding. This matrix includes: Physical feature calculations are performed based on the raw data. Based on the latitude and longitude of the power station and the timestamp, the solar position features are determined, including the solar altitude angle and the solar azimuth angle. Based on the aforementioned solar position characteristics, the theoretical clear-sky irradiance is determined; Based on the theoretical clear-sky irradiance, the clear-sky index is determined; The historical enhanced meteorological matrix is ​​constructed by splicing together the historical meteorological observation sequence, the solar position characteristics, the theoretical clear sky irradiance, the clear sky index, and the time period code. Based on the meteorological forecasting system, the future meteorological forecast sequence within the forecast time domain is determined, and based on the future meteorological forecast sequence and the theoretical clear sky irradiance, the predicted clear sky index is determined. The future weather matrix is ​​constructed by splicing together the solar position characteristics, the theoretical clear sky irradiance, the predicted clear sky index, and the time period code.

[0026] In some exemplary embodiments, quality control is performed before physical feature calculations are performed on the raw data. The quality control includes outlier removal, missing data imputation, and noise interference elimination.

[0027] As an example, quality control based on the raw data is performed as follows: abnormal records of non-zero power and exceeding rated capacity at night are removed; missing values ​​are imputed based on weather conditions, with the difference from the clear sky model used for sunny days and the mean of the adjacent time period used for cloudy and rainy days; during nighttime periods (solar altitude angle)... Power is reduced to zero to eliminate nighttime noise interference.

[0028] As an example, after quality control is completed, material characteristics are calculated. Based on the power station's latitude and longitude, Julian day, and true solar time, the solar altitude angle is calculated using astronomical formulas. and solar azimuth Then calculate the theoretical clear-sky irradiance. : ; in, Indicates theoretical clear-sky irradiance; Represents the solar constant; Indicates Julian Japan; Indicates the solar altitude angle; It represents atmospheric transmittance.

[0029] The theoretical clear-sky irradiance in the above formula describes the maximum irradiance reaching the ground under conditions of complete atmospheric transparency. It also serves as a unified benchmark for calculating the clear-sky index and normalizing subsequent weather forecast deviations.

[0030] As an example, the clear sky index is calculated based on theoretical clear sky irradiance. : ; in, Indicates the clear sky index; Indicates the measured irradiance; Indicates theoretical clear-sky irradiance; This indicates the prevention of zero small quantities.

[0031] The physical meaning of the clear sky index in the above formula is the overall degree of cloud cover on irradiance, which is a core intermediate quantity for characterizing weather conditions.

[0032] As an example, after completing the physical feature calculation, the historical meteorological observation sequence, solar position characteristics, theoretical clear sky irradiance, clear sky index, and time period encoding are concatenated to construct the historical enhanced meteorological matrix. (dimension) The time cycle encoding uses a sine-cosine cyclic form to preserve the phase continuity of the daily and annual cycles. ; in, Represents a historical enhanced meteorological matrix; Indicates the length of the matrix; This represents the characteristic dimensions of the historical enhanced meteorological matrix; Indicates total horizontal radiation; Indicates ambient temperature; Indicates wind speed; Indicates relative humidity; Indicates the solar altitude angle; Indicates the solar azimuth angle; This indicates the clear sky index.

[0033] As an example, referencing historical enhanced meteorological matrices, the forecast time domain is obtained from the meteorological forecasting system. Construct a future weather matrix based on the future weather forecast sequence within the range. : ; in, Indicates future meteorological dimensions, dimension , Indicates the length of the prediction time domain, and indicates the total length of the prediction time domain. One time step; Characteristic dimensions representing future weather patterns; This indicates the predicted irradiance (based on irradiance forecasts obtained from meteorological forecasting systems). Indicates the predicted temperature; This indicates the predicted cloud cover (based on cloud cover forecasts obtained from meteorological forecasting systems). Indicates the solar altitude angle; Indicates the solar azimuth angle; This indicates a predicted clear sky index.

[0034] In the above formula, , , The deterministic physical quantities are precisely calculated using astronomical formulas and do not include forecast errors. The future meteorological matrix will encode the deterministic physical quantities separately from the forecast quantities that include errors, so that the model can distinguish the uncertainty differences between the two types of information at the input end, providing a normalization basis for subsequent quality assessment.

[0035] Step S103: Based on the deviation between the meteorological forecast value and the actual meteorological value, generate a credibility score and dynamic fusion weight.

[0036] In some exemplary embodiments, a confidence score and dynamic fusion weight are generated based on the deviation between the weather forecast value and the observed weather value, including: Based on the theoretical clear-sky irradiance, the deviation between the weather forecast value and the actual weather measurement value at each time point within the historical window is determined; Based on the aforementioned deviation, the overall error is determined by exponential decay weighting. Based on the comprehensive error, the reliability score of the current weather forecast is determined through exponential mapping transformation; The confidence score, the comprehensive error, and the predicted clear sky index are concatenated and input into a gating network to generate dynamic fusion weights.

[0037] In some exemplary embodiments, the credibility score is used for the comprehensive screening of subsequent meteorological block embedding sequences, and the dynamically fused weights are used for the subsequent cross-attention mechanism.

[0038] In some exemplary embodiments, based on historical enhanced meteorological matrices (dimension) ), determine the history window ( (a moment).

[0039] As an example, based on history window Calculate the deviation between verifiable weather forecasts and observed weather values ​​at each time point: ; in, Indicates the first Deviation at any given moment; express Measured irradiance at any given time; express Predicted irradiance at any given time; express Theoretical clear-sky irradiance at any given moment; In the above formula, with Instead As a normalization denominator, it can avoid the influence of strong cloud cover during certain periods. The numerical values ​​are too low, which leads to an overestimation of the bias, thus excessively penalizing the ability of numerical weather low irradiance forecasting.

[0040] As an example, the overall error is determined based on the bias using exponentially decaying weighting: ; in, Indicates the overall error; Indicates the attenuation coefficient; Indicates the first Deviation at a given moment.

[0041] The overall error in the above formula is used to assign a reference weight to the deviation. The smaller the value, the stronger the dominant role of the recent time deviation, which is suitable for scenarios with drastic weather changes; The larger the value, the more evenly the contribution of historical time deviations is distributed, which is suitable for scenarios with stable weather.

[0042] As an example, based on the comprehensive error, the reliability score of the current weather forecast is determined through exponential mapping transformation: ; in, Indicates credibility score; Indicates the overall error; This represents the sensitivity coefficient.

[0043] In the above formula, The closer the value is to 1, the smaller the error in the numerical weather forecast and the higher its reliability. When the value approaches 0, it indicates that the numerical weather forecast has a large error in the near term and is unreliable. The model automatically degenerates into a pure historical data-driven mode, thus systematically suppressing low-quality weather forecasts.

[0044] As an example, the confidence score, comprehensive error, and predicted clear sky index are concatenated and input into a gating network to generate dynamic fusion weights: ; in, Indicates dynamic fusion weights; Indicates the activation function; , The learnable parameters of a lightweight linear gating unit; Indicates credibility score; Indicates the overall error; This represents the mean of the predicted clear sky index within the prediction time domain; Indicates vector transpose; This represents the preset base fusion weights.

[0045] The dynamic fusion weights in the above formula are used to control the weights of meteorological blocks participating in the cross-attention mechanism.

[0046] Step S104: Decompose the historical photovoltaic power sequence to determine the trend component and fluctuation component. Obtain the power block embedding sequence by power dual-component block embedding. Perform block embedding on the meteorological matrix to obtain the meteorological block embedding sequence.

[0047] In some exemplary embodiments, the historical photovoltaic power sequence is decomposed to determine trend components and fluctuation components. A power block embedding sequence is obtained through power dual-component block embedding. The meteorological matrix is ​​then block-embedded to obtain a meteorological block embedding sequence, including: Based on the historical photovoltaic power sequence, the low-frequency trend component and high-frequency fluctuation component are determined by exponential moving average decomposition. Based on the historical enhanced meteorological matrix and the future meteorological matrix, the corresponding trend component and fluctuation component are determined through the exponential moving average decomposition. The corresponding trend component and fluctuation component are respectively divided into blocks along the time axis to obtain a block sequence; Based on the block sequence, the block is embedded by mapping to a latent space of a preset dimension through a linear layer, and a positional encoding corresponding to the original time position of the block sequence is applied to obtain a meteorological block embedding sequence. The low-frequency trend component and the high-frequency fluctuation component are respectively divided into blocks to obtain their respective power block sequences. The corresponding power block embedding sequence is obtained by power dual-component block embedding.

[0048] In some exemplary embodiments, the photovoltaic power sequence simultaneously includes a low-frequency trend component driven by the gradual change of the sun's position and a high-frequency fluctuation component generated by random cloud shading. The two types of components have different physical origins and frequency structures, and modeling them simultaneously by a single network will cause mutual interference in the learning process.

[0049] In some exemplary embodiments, the historical photovoltaic power sequence is input. Historical Enhanced Weather Matrix Future Weather Matrix Credibility score and dynamic fusion weights Predicting the future Step power sequence .

[0050] As an example, an exponential moving average decomposition is performed on the historical photovoltaic power series to identify low-frequency trend components and high-frequency fluctuation components: ; in, Indicates low-frequency trend components; Indicates high-frequency fluctuation components; Indicates an exponential moving average; This represents the smoothing coefficient.

[0051] As an example, the historical enhanced weather matrix and future weather matrix Performing the same exponential moving average decomposition, the historical enhanced meteorological matrices were obtained respectively. Corresponding trend components and fluctuation components Future Weather Matrix Corresponding trend components and fluctuation components Then, normalization is performed on each component to eliminate batch-to-batch distribution drift: ; in, The normalized component represents the normalized component. Represents any component to be normalized ( , , , , , ); Representing components The mean of the current sample; Representing components Standard deviation on the current sample; This indicates the prevention of zero small quantities.

[0052] In the above formula, the true labels of each component are obtained by performing the same moving average decomposition on the target sequence, which is used to maintain decomposition consistency.

[0053] As an example, historical enhanced weather matrix Each component , and future weather matrix Each component , By block length along the time axis The sequence is divided into non-overlapping blocks and then mapped through a linear layer to... In the latent space, block embedding is completed; taking the historical enhanced meteorological matrix as an example: ; in, This indicates the result after block splitting. A sequence of non-overlapping blocks, with dimension . ; Represents a historical enhanced meteorological matrix; This represents the total number of time steps in the historical enhanced meteorological matrix; The number of blocks in the sequence, rounded down; Indicates the block length; Indicates block segmentation; Represents a linear layer mapping; Represents the dimension of the latent space; This represents the initial feature sequence after block embedding, with dimension . .

[0054] In some exemplary embodiments, after block embedding is completed, invalid blocks are removed through a two-stage filtering process. The first stage is physical constraint hard filtering: removing invalid blocks from the block sequence based on the solar altitude angle. The first stage involves removing blocks; the second stage involves performing credibility-linked filtering on the remaining block sequences after the first stage: calculating a comprehensive score for each remaining block sequence, taking the preset number of blocks with the highest comprehensive score, and completing the final embedding through a trainable linear layer to obtain the meteorological block embedding sequence, while retaining the original time and location information.

[0055] As an example, the first stage removes nighttime blocks (solar altitude angle). ), and the remaining daytime block (solar altitude angle) ) Create an index set Number of daytime blocks Calculate a comprehensive score for each daytime block: ; in, Indicates the first The overall score of each daytime block; Indicates credibility score; This represents a two-layer linear scoring network for meteorological embedding blocks. Learnable rating; Indicates power embedding block The information content score is determined by the internal variance.

[0056] In the above formula, if the score approaches 1, the overall score is dominated by the meteorological embedding block; if the score approaches 0, the overall score is dominated by the power embedding block, which is used to avoid the interference of low-quality weather forecasts on the two-stage screening.

[0057] As an example, the one with the highest overall score is selected. Each daytime block is embedded using a trainable linear layer to obtain a weather block embedding sequence, while retaining the original time and location information: ; in, This indicates that the highest overall score will be selected. A set of indices for daytime blocks; This indicates that the highest overall score will be selected. A daytime block; Indicates the first The overall score of each daytime block; This represents the final weather patch embedding sequence, with dimensions... ; Indicates a linear projection layer; This indicates the position code.

[0058] In the above formula, the original time and location information is preserved based on dynamic location recoding; the meteorological block embedding sequence is used for subsequent cross-attention fusion.

[0059] As an example, the power two-component block embedding could be a pair of power trend components. and power fluctuation components Perform block partitioning separately to obtain the corresponding power block sequences. , Based on the physical characteristics of the trend component and the fluctuation component, a differentiated normalization embedding strategy is adopted to perform normalization processing on the power block sequence corresponding to the trend component, and then map it to a linear layer. In the latent space, the original amplitude of the power block sequence corresponding to the fluctuation component is preserved, and then mapped to a linear layer. By using the hidden space, we obtain the corresponding power block embedding sequence: ; in, Indicates the first The power block sequence corresponding to each trend component; Indicates the first The power block sequence corresponding to each fluctuation component; express The mean; express Standard deviation; Indicates a linear projection layer; express The embedding vector; express The embedding vector; This represents the power block embedding sequence corresponding to the trend component, with dimensions... ; This represents the power block embedding sequence corresponding to the fluctuation component.

[0060] In the above formula, mean-variance normalization is performed on the power block sequence corresponding to the trend component to eliminate inter-block amplitude drift; the original amplitude is retained for the power block sequence corresponding to the fluctuation component, and mapped to a linear layer. Hidden space, which protects the phase and amplitude structure of wave components.

[0061] Step S105: Based on the first and second neural networks, perform heterogeneous dual-stream parallel modeling on the power block embedding sequence to obtain trend branch and fluctuation branch. Based on the dynamic fusion weight and cross attention, fuse the meteorological block embedding sequence into the two branches.

[0062] In some exemplary embodiments, heterogeneous dual-stream parallel modeling is performed on the power block embedding sequence based on the first and second neural networks to obtain a trend branch and a fluctuation branch. Based on the dynamic fusion weights and cross attention, the meteorological block embedding sequence is fused into the two branches, including: Based on the first neural network, the power block embedding sequence corresponding to the low-frequency trend component is processed to obtain the trend branch; Based on the second neural network, the power block embedding sequence corresponding to the high-frequency fluctuation component is processed to obtain the fluctuation branch; Based on the fluctuation branch, a query is performed, using the historical meteorological block embedding sequence and the future meteorological block embedding sequence as keys respectively. Two independent multi-head cross attention is executed, and the future meteorological block embedding sequence is adaptively controlled based on the dynamic fusion weight. The fused fluctuation branch is obtained through the first neural network. The multi-head cross attention is replaced by a lightweight linear weighted fusion to obtain the fused trend branch.

[0063] In some exemplary embodiments, the first neural network may be a feedforward network, and the second neural network may be a multi-head self-attention network.

[0064] In some exemplary embodiments, the trend component is a low-frequency smooth signal with strong linear predictability. However, it is prone to overfitting based on a complex attention mechanism. Therefore, the trend branch is obtained by modeling based on a two-layer feedforward network.

[0065] In some exemplary embodiments, the meteorological block corresponding to the trend component is embedded in the sequence, and after linear projection, it is added to the residual of the trend branch to introduce meteorological auxiliary information into the trend branch, so that the trend branch combines the meteorological auxiliary information to predict photovoltaic power.

[0066] In some exemplary embodiments, the fluctuation component has a strong periodic structure and nonlinear perturbation, and global attention is required to capture the nonlinear dependence. Therefore, the fluctuation branch is obtained based on multi-head self-attention modeling.

[0067] In some exemplary embodiments, historical meteorological observation information and future meteorological information are respectively incorporated into the power representation, and the fusion weight of future meteorological information is controlled by dynamic fusion weights; taking the fluctuation branch as an example, using For querying, historical weather block embedding sequences are used respectively. and future weather block embedding sequences Using the key value, perform two independent multi-head cross-attention operations to obtain the fused fluctuation branch.

[0068] As an example, based on feedforward network modeling, the power block embedding sequence corresponding to the trend component is processed to obtain the trend branch; the meteorological block embedding sequence corresponding to the trend component is linearly projected and added to the residual of the trend branch to introduce meteorological auxiliary information into the trend branch. ; in, This represents the intermediate quantity obtained by adding the residuals of the feedforward network and the trend branch; Presentation layer normalization processing; This represents the power block embedding sequence corresponding to the trend component; This represents a two-layer feedforward network; Indicates trend branches, dimensions ; This represents the meteorological block embedding sequence corresponding to the trend component; Indicates a linear projection layer; This indicates that meteorological auxiliary information will be injected into the trend branch.

[0069] As an example, meteorological auxiliary information can be a meteorological block embedding sequence obtained by block embedding of historical enhanced meteorological matrices and future meteorological matrices.

[0070] As an example, based on multi-head self-attention modeling, the power block embedding sequence corresponding to the fluctuation component is processed to obtain the fluctuation branch: ; in, express The trend; This represents the power block embedding sequence corresponding to the fluctuation component; This indicates an average pooling operation, used to extract... Local trends; Indicates the boundary fill operation; Indicates the fluctuation component after trend removal; Presentation layer normalization processing; This indicates a multi-head self-attention mechanism; This represents the intermediate quantity obtained after layer normalization and multi-head self-attention processing; Indicates a feedforward network; Represents the fluctuating branch, dimension .

[0071] In the above formula, to suppress spurious attention caused by short-term non-stationarity between power blocks, the following is applied: Perform detrending processing first, then use it as a query key, preserving the original... .

[0072] As an example, the historical weather block embedding sequence is used respectively. and future weather block embedding sequences Using the key value, perform two independent multi-head cross-attention operations to obtain the fused fluctuation branch: ; in, Represents the embedded sequence of historical weather blocks Cross-attention output; Indicates future weather block embedding sequence Cross-attention output; , , This represents the learnable projection matrix. Used to map input to query. Used to map input to keys, Used to map inputs to values; This represents the intermediate quantity after the merging of fluctuating branches; This represents the fused fluctuation branch; This represents a feedforward network.

[0073] In the above formula, lightweight linear weighted fusion replaces bullish attention and processes the trend branches to obtain the fused trend branches.

[0074] Step S106: Calculate and determine the current weather state based on the physical characteristics, adaptively generate the dynamic weights of the two branches, obtain the predicted values ​​of each branch and weighted fuse them, and output the future photovoltaic power prediction sequence.

[0075] In some exemplary embodiments, the current weather state is calculated and determined based on the physical characteristics, the dynamic weights of the two branches are adaptively generated, the predicted values ​​of each branch are obtained and weighted and fused, and the future photovoltaic power prediction sequence is output, including: The current weather state is determined based on the physical characteristics. The weather state feature vector is determined based on the clear sky index, the average cloud cover forecast in the prediction time domain, and the confidence score. The weather state includes sunny, cloudy, and rainy days. The weather state feature vector is input into a linear weight generation network, and normalized by a normalized exponential function to obtain the dynamic weights of the fused trend branch and the dynamic weights of the fused fluctuation branch. A linear decoder is applied to the fused trend branch and the fused fluctuation branch respectively to obtain the predicted values ​​of the trend branch and the fluctuation branch. Based on the dynamic weights of the fused trend branch and the fused fluctuation branch, a weighted fusion is performed to obtain the future photovoltaic power prediction sequence.

[0076] In some exemplary embodiments, the weather is classified into sunny days, cloudy days, and rainy days based on the clear sky index and coefficient of variation calculated from physical characteristics.

[0077] As an example, based on the clear sky index and coefficient of variation calculated from physical characteristics, the weather is divided into sunny days, cloudy days, and rainy days. Among these, if the daily average clear sky index is... And coefficient of variation If the average daily clear sky index is [value missing], then the weather is classified as sunny; if the average daily clear sky index is [value missing], then the weather is classified as sunny. or coefficient of variation If the average daily clear sky index is [value missing], then the weather will be classified as cloudy; if the average daily clear sky index is [value missing], then the weather will be classified as cloudy. Then the weather is classified as cloudy or rainy.

[0078] As an example, a weather state feature vector is determined based on the clear sky index, the mean cloud cover forecast in the prediction time domain, and the confidence score: ; in, Represents the feature vector of weather conditions; This represents the average clear sky index within a historical window. The standard deviation of the clear sky index within a historical window; This represents the average cloud cover forecast over the predicted time domain; This indicates the credibility score.

[0079] As an example, weather state feature vectors are input into a linear weighted generation network, and then... After normalization, the dynamic weights of the merged trend branch and the merged fluctuation branch are obtained: ; in, Represents the overall dynamic weight; This represents the dynamic weight of the merged trend branches; This represents the dynamic weight of the fluctuating branch after fusion; , This represents the learnable parameters.

[0080] In the above formula, the linear weight generation network automatically learns the mapping between weather conditions and branch weights during training. For example, a sunny day (…). high, When the value is low, the dynamic weight of the merged trend branch The trend is increasing, and at this point, the trend branch dominates; cloudy weather with sudden changes ( high, When the dynamic weights of the merged fluctuating branches are low, When the trend is large, the fluctuating branch dominates; during cloudy or rainy weather, the two branches tend to be in equilibrium.

[0081] As an example, a linear decoder is applied to the merged trend branch and the merged fluctuation branch respectively. The predicted values ​​of each branch are then weighted and merged to obtain the future photovoltaic power prediction sequence: ; in, This represents the predicted value of the trend branch; This represents the predicted value of the fluctuating branch; Indicates a linear decoding layer; This represents a sequence of future photovoltaic power forecasts.

[0082] In some exemplary embodiments, the training process of the first neural network and the second neural network described above uses a time-domain-frequency-domain joint loss function to calculate the loss for the trend branch and the fluctuation branch respectively, and then determines the total loss.

[0083] As an example, the joint time-frequency domain loss function: ; in, Indicates the total loss; Indicates the loss along the trend branch; This indicates the loss of the fluctuating branch; Indicates time-domain loss, superscript for Time represents the time-domain loss of the trend branch, superscript for Time represents the time-domain loss of the fluctuating branch; Indicates the branch road number The predicted value of the step; Indicates the branch road number The true label of the step; Indicates frequency domain loss, superscript for Time represents the frequency domain loss of the trend branch, superscript for Time represents the frequency domain loss of the fluctuating branch; This represents the total number of steps in the prediction time domain; Indicates Fast Fourier Transform; This indicates that taking the complex modulus yields the amplitude spectrum; This represents the frequency domain weight.

[0084] In the above formula, the time-domain term uses the mean absolute error to measure the progressive prediction deviation and constrain the time-domain accuracy; the frequency-domain term performs a fast Fourier transform on the predicted sequence and the true sequence respectively, takes the complex modulus, calculates the amplitude spectrum, explicitly constrains the daily periodic spectrum structure, and prevents spectral distortion caused by a single time-domain loss; the loss is calculated for each branch separately to eliminate the error complementarity effect between branches and ensure that the network parameters of each branch are optimized for the component it is responsible for.

[0085] The beneficial effects of this disclosure are: This disclosure performs quality control and physical characteristic calculations based on raw data to determine solar position characteristics and theoretical clear-sky irradiance, realizing physical prior constraints and optimizing redundant ineffective meteorological features.

[0086] This disclosure optimizes the quality of using future meteorological information by generating a credibility score and dynamic fusion weights through bias, takes into account the forecast quality of future meteorological information, and improves the prediction stability and accuracy of the model under different weather conditions.

[0087] This disclosure uses weather-sensing heterogeneous dual-flow decoupling prediction to decompose the photovoltaic power sequence into trend components and fluctuation components, and uses a first neural network and a second neural network for differentiated processing, which effectively avoids mutual interference caused by the mixed modeling of two types of signals with different physical causes and frequency structures, and improves the prediction accuracy of the model under different weather conditions.

[0088] This disclosure calculates and classifies weather based on physical characteristics, determines weather state feature vectors, and obtains the dynamic weights of the fused trend branch and the dynamic weights of the fused fluctuation branch through normalization processing using a normalized exponential function. This effectively enhances trend modeling under sunny weather conditions and enhances fluctuation modeling under cloudy and abrupt weather conditions, enabling the model to automatically adjust its modeling focus according to weather changes, thereby improving the model's predictive consistency and predictive adaptability across weather types.

[0089] To implement the above exemplary embodiments, this disclosure also proposes a weather-sensing heterogeneous dual-flow decoupling prediction device based on deep learning.

[0090] Figure 3 This is a schematic diagram of a deep learning-based weather sensing heterogeneous dual-flow decoupling forecasting device 200, provided as an exemplary embodiment of this disclosure. The device 200 can be implemented by software and / or hardware and is generally integrated into an electronic device. For example... Figure 3 As shown, the device 200 includes: a determining module 201, a constructing module 202, a generating module 203, an embedding module 204, a fusion module 205, and an output module 206, wherein, The determination module 201 is configured to determine the raw data in the prediction time domain, the raw data including historical photovoltaic power sequences, historical meteorological observation sequences, future meteorological forecast sequences, power station latitude and longitude, and timestamps; The construction module 202 is configured to perform physical feature calculations based on the original data, determine the solar position features and theoretical clear sky irradiance, and construct a meteorological matrix by splicing together the historical meteorological observation sequence, the solar position features, the theoretical clear sky irradiance and the time period code. The generation module 203 is configured to generate a credibility score and dynamic fusion weights based on the deviation between the meteorological forecast value and the actual meteorological value. The embedding module 204 is configured to decompose the historical photovoltaic power sequence to determine the trend component and fluctuation component, obtain the power block embedding sequence through power dual-component block embedding, and perform block embedding on the meteorological matrix to obtain the meteorological block embedding sequence. The fusion module 205 is configured to perform heterogeneous dual-stream parallel modeling on the power block embedding sequence based on the first and second neural networks to obtain trend branches and fluctuation branches, and to fuse the meteorological block embedding sequence into the two branches based on the dynamic fusion weights and cross attention. The output module 206 is configured to calculate and determine the current weather state based on the physical characteristics, adaptively generate the dynamic weights of the two branches, obtain the predicted values ​​of each branch and weighted fuse them, and output the future photovoltaic power prediction sequence.

[0091] The deep learning-based weather perception heterogeneous dual-flow decoupling prediction device provided in the exemplary embodiments of this disclosure can execute the deep learning-based weather perception heterogeneous dual-flow decoupling prediction method provided in any embodiment of this disclosure, and has the corresponding functional modules and beneficial effects of the method execution.

[0092] To implement the above embodiments, this disclosure also proposes a computer program product, including a computer program / instruction, which, when executed by a processor, implements the deep learning-based weather perception heterogeneous dual-flow decoupling prediction method in the above embodiments.

[0093] Figure 4 A schematic diagram of the structure of an electronic device provided for an exemplary embodiment of this disclosure.

[0094] The following is a detailed reference. Figure 4 The diagram illustrates a structural schematic suitable for implementing the electronic device 300 in the exemplary embodiments of this disclosure. The electronic device 300 in the exemplary embodiments of this disclosure may include, but is not limited to, mobile terminals such as mobile phones, laptops, digital broadcast receivers, PDAs (personal digital assistants), PADs (tablet computers), PMPs (portable multimedia players), in-vehicle terminals (e.g., in-vehicle navigation terminals), and fixed terminals such as digital TVs and desktop computers. Figure 4 The electronic device shown is merely an example and should not impose any limitation on the functionality and scope of use of the exemplary embodiments disclosed herein.

[0095] like Figure 4As shown, the electronic device 300 may include a processor (e.g., a central processing unit, a graphics processing unit, etc.) 301, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 302 or a program loaded from a memory 308 into a random access memory (RAM) 303. The RAM 303 also stores various programs and data required for the operation of the electronic device 300. The processor 301, ROM 302, and RAM 303 are interconnected via a bus 304. An input / output (I / O) interface 305 is also connected to the bus 304.

[0096] Typically, the following devices can be connected to I / O interface 305: input devices 306 including, for example, touchscreens, touchpads, keyboards, mice, cameras, microphones, accelerometers, gyroscopes, etc.; output devices 307 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; memory devices 308 including, for example, magnetic tapes, hard disks, etc.; and communication devices 309. Communication device 309 allows electronic device 300 to communicate wirelessly or wiredly with other devices to exchange data. Although Figure 4 An electronic device 300 with various devices is shown; however, it should be understood that it is not required to implement or possess all of the devices shown. More or fewer devices may be implemented or possessed alternatively.

[0097] In particular, according to embodiments of this disclosure, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this disclosure include a computer program product comprising a computer program carried on a non-transitory computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication device 309, or installed from memory 308, or installed from ROM 302. When the computer program is executed by processor 301, it performs the functions defined in the deep learning-based weather sensing heterogeneous dual-flow decoupling prediction method of exemplary embodiments of this disclosure.

[0098] It should be noted that the computer-readable medium described in this disclosure can be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. A computer-readable storage medium can be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this disclosure, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In this disclosure, a computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium can be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wires, optical fibers, RF (radio frequency), etc., or any suitable combination thereof.

[0099] In some implementations, clients and servers can communicate using any currently known or future-developed network protocol such as HTTP (Hypertext Transfer Protocol) and can interconnect with digital data communication (e.g., communication networks) of any form or medium. Examples of communication networks include local area networks (“LANs”), wide area networks (“WANs”), the Internet (e.g., the Internet of Things), and peer-to-peer networks (e.g., ad hoc peer-to-peer networks), as well as any currently known or future-developed networks.

[0100] The aforementioned computer-readable medium may be included in the aforementioned electronic device; or it may exist independently and not assembled into the electronic device.

[0101] The aforementioned computer-readable medium carries one or more programs, which, when executed by the electronic device, cause the electronic device to perform the aforementioned deep learning-based weather perception heterogeneous dual-flow decoupling prediction method.

[0102] Electronic devices can be programmed with computer program code in one or more programming languages ​​or combinations thereof to perform the operations of this disclosure. These programming languages ​​include, but are not limited to, object-oriented programming languages ​​such as Java, Smalltalk, and C++, as well as conventional procedural programming languages ​​such as "C" or similar languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0103] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0104] The units described in the exemplary embodiments of this disclosure can be implemented in software or in hardware. The names of the units are not, in some cases, intended to limit the specific unit.

[0105] The functions described above in this document can be performed at least in part by one or more hardware logic components. For example, exemplary types of hardware logic components that can be used, without limitation, include: field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), system-on-a-chip (SoCs), complex programmable logic devices (CPLDs), and so on.

[0106] In the context of this disclosure, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0107] The above description is merely a preferred embodiment of this disclosure and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of this disclosure is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-described concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features disclosed in this disclosure that have similar functions.

[0108] Furthermore, while the operations are described in a specific order, this should not be construed as requiring these operations to be performed in the specific order shown or in a sequential order. In certain environments, multitasking and parallel processing may be advantageous. Similarly, while several specific implementation details are included in the above discussion, these should not be construed as limiting the scope of this disclosure. Certain features described in the context of individual embodiments may also be implemented in combination in a single embodiment. Conversely, various features described in the context of a single embodiment may also be implemented individually or in any suitable sub-combination in multiple embodiments.

[0109] Although the subject matter has been described using language specific to structural features and / or methodological logic, it should be understood that the subject matter defined in the appended claims is not necessarily limited to the specific features or actions described above. Rather, the specific features and actions described above are merely illustrative examples of implementing the claims.

Claims

1. A weather-aware heterogeneous dual-flow decoupling prediction method based on deep learning, characterized in that, include: The original data for the prediction time domain is determined, including historical photovoltaic power sequences, historical meteorological observation sequences, future meteorological forecast sequences, power plant latitude and longitude, and timestamps; Based on the original data, physical features are calculated to determine the solar position features and theoretical clear-sky irradiance. The historical meteorological observation sequence, the solar position features, the theoretical clear-sky irradiance, and the time period code are then spliced ​​together to construct a meteorological matrix. Based on the deviation between meteorological forecasts and actual meteorological values, a reliability score and dynamic fusion weights are generated. Based on the historical photovoltaic power sequence, the trend component and fluctuation component are determined. Through power dual-component block embedding, a power block embedding sequence is obtained. The meteorological matrix is ​​then block-embedded to obtain a meteorological block embedding sequence. Heterogeneous dual-stream parallel modeling is performed on the power block embedding sequence based on the first and second neural networks to obtain trend branches and fluctuation branches. Based on the dynamic fusion weights and cross attention, the meteorological block embedding sequence is fused into the two branches. Based on the physical characteristics, the current weather condition is calculated and determined. The dynamic weights of the two branches are adaptively generated, the predicted values ​​of each branch are obtained and weighted and fused, and the future photovoltaic power prediction sequence is output.

2. The method of claim 1, wherein, The historical meteorological observation sequence includes at least total horizontal radiation, ambient temperature, wind speed, and relative humidity; The timestamps include Julian Day and True Solar Time; The meteorological matrix includes a historical enhanced meteorological matrix and a future meteorological matrix; Based on the raw data, physical feature calculations are performed to determine solar position characteristics and theoretical clear-sky irradiance. A meteorological matrix is ​​constructed by concatenating the historical meteorological observation sequence, the solar position characteristics, the theoretical clear-sky irradiance, and the time period encoding. This matrix includes: Physical feature calculations are performed based on the raw data. Based on the latitude and longitude of the power station and the timestamp, the solar position features are determined, including the solar altitude angle and the solar azimuth angle. Based on the aforementioned solar position characteristics, the theoretical clear-sky irradiance is determined; Based on the theoretical clear-sky irradiance, the clear-sky index is determined; The historical enhanced meteorological matrix is ​​constructed by splicing together the historical meteorological observation sequence, the solar position characteristics, the theoretical clear sky irradiance, the clear sky index, and the time period code. Based on the meteorological forecasting system, the future meteorological forecast sequence within the forecast time domain is determined, and based on the future meteorological forecast sequence and the theoretical clear sky irradiance, the predicted clear sky index is determined. The future weather matrix is ​​constructed by splicing together the solar position characteristics, the theoretical clear sky irradiance, the predicted clear sky index, and the time period code.

3. The method of claim 1, wherein, Based on the deviation between meteorological forecasts and actual meteorological values, a reliability score and dynamic fusion weights are generated, including: Based on the theoretical clear-sky irradiance, the deviation between the weather forecast value and the actual weather measurement value at each time point within the historical window is determined; Based on the aforementioned deviation, the overall error is determined by exponential decay weighting. Based on the comprehensive error, the reliability score of the current weather forecast is determined through exponential mapping transformation; The confidence score, the comprehensive error, and the predicted clear sky index are concatenated and input into a gating network to generate dynamic fusion weights.

4. The method of claim 1, wherein, Based on the historical photovoltaic power sequence, trend components and fluctuation components are determined. A power block embedding sequence is obtained through power dual-component block embedding. The meteorological matrix is ​​then subjected to block embedding to obtain a meteorological block embedding sequence, including: Based on the historical photovoltaic power sequence, the low-frequency trend component and high-frequency fluctuation component are determined by exponential moving average decomposition. Based on the historical enhanced meteorological matrix and the future meteorological matrix, the corresponding trend component and fluctuation component are determined through the exponential moving average decomposition. The corresponding trend component and fluctuation component are respectively divided into blocks along the time axis to obtain a block sequence; Based on the block sequence, the block is embedded by mapping to a latent space of a preset dimension through a linear layer, and a positional encoding corresponding to the original time position of the block sequence is applied to obtain a meteorological block embedding sequence. The low-frequency trend component and the high-frequency fluctuation component are respectively divided into blocks to obtain their respective power block sequences. The corresponding power block embedding sequence is obtained by power dual-component block embedding.

5. The method of claim 1, wherein, The two branches include trend branches and fluctuation branches; The heterogeneous dual-stream parallel modeling of the power block embedding sequence based on the first and second neural networks yields a trend branch and a fluctuation branch. Based on the dynamic fusion weights and cross-attention, the meteorological block embedding sequence is fused into the two branches, including: Based on the first neural network, the power block embedding sequence corresponding to the low-frequency trend component is processed to obtain the trend branch; Based on the second neural network, the power block embedding sequence corresponding to the high-frequency fluctuation component is processed to obtain the fluctuation branch; Based on the fluctuation branch, a query is performed, using the historical meteorological block embedding sequence and the future meteorological block embedding sequence as keys respectively. Two independent multi-head cross attention is executed, and the future meteorological block embedding sequence is adaptively controlled based on the dynamic fusion weight. The fused fluctuation branch is obtained through the first neural network. The multi-head cross attention is replaced by a lightweight linear weighted fusion to obtain the fused trend branch.

6. The method according to claim 1, characterized in that, The predicted values ​​for each branch include the predicted values ​​for trend branches and the predicted values ​​for fluctuation branches; The process involves calculating and determining the current weather state based on the physical characteristics, adaptively generating dynamic weights for the two branches, obtaining predicted values ​​for each branch, weighted and fused, and outputting a future photovoltaic power prediction sequence, including: The current weather state is determined based on the physical characteristics. The weather state feature vector is determined based on the clear sky index, the average cloud cover forecast in the prediction time domain, and the confidence score. The weather state includes sunny, cloudy, and rainy days. The weather state feature vector is input into a linear weight generation network, and normalized by a normalized exponential function to obtain the dynamic weights of the fused trend branch and the dynamic weights of the fused fluctuation branch. A linear decoder is applied to the fused trend branch and the fused fluctuation branch respectively to obtain the predicted values ​​of the trend branch and the fluctuation branch. Based on the dynamic weights of the fused trend branch and the fused fluctuation branch, a weighted fusion is performed to obtain the future photovoltaic power prediction sequence.

7. A deep learning-based weather sensing heterogeneous dual-flow decoupling prediction device, the device comprising: The determination module is configured to determine the raw data in the prediction time domain, which includes historical photovoltaic power sequences, historical meteorological observation sequences, future meteorological forecast sequences, power plant latitude and longitude, and timestamps; The construction module is configured to perform physical feature calculations based on the original data, determine the solar position features and theoretical clear sky irradiance, and construct a meteorological matrix by splicing together the historical meteorological observation sequence, the solar position features, the theoretical clear sky irradiance, and the time period code. The generation module is configured to generate a credibility score and dynamic fusion weights based on the deviation between the weather forecast and the actual weather value. The embedding module is configured to decompose the historical photovoltaic power sequence to determine the trend component and fluctuation component, obtain the power block embedding sequence through power dual-component block embedding, and perform block embedding on the meteorological matrix to obtain the meteorological block embedding sequence. The fusion module is configured to perform heterogeneous dual-stream parallel modeling on the power block embedding sequence based on the first and second neural networks to obtain trend branches and fluctuation branches, and to fuse the meteorological block embedding sequence into the two branches based on the dynamic fusion weights and cross attention. The output module is configured to calculate and determine the current weather state based on the physical characteristics, adaptively generate the dynamic weights of the two branches, obtain the predicted values ​​of each branch and weighted fuse them, and output the future photovoltaic power prediction sequence.

8. An electronic device, comprising: include: Memory; processor; as well as Computer programs; The computer program is stored in the memory and configured to be executed by the processor to implement the steps of the method as described in any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, It stores a computer program / instruction thereon, which, when executed by a processor, implements the steps of the method described in any one of claims 1-6.

10. A computer program product comprising computer programs / instructions, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method according to any one of claims 1-6.

Images