A solar irradiance prediction method fusing physical prior and multi-scale features

Abstract

The present application relates to a kind of solar irradiance prediction methods of fusing physical priori and multiscale feature, first, each meteorological type data is divided into fast variable space and slow variable space, and introduce solar zenith angle as certainty benchmark, simultaneously, the trend item and detail item of surface solar irradiance sequence are extracted by adaptive signal decomposition;Then construct to be trained network, first, with trend modeling branch and detail modeling branch maintain linear calculation complexity, while, respectively, extract deep trend feature map and deep detail feature map, then join step perception dynamic fusion decoding module, according to the prediction lead time adaptive adjustment double-branch weight, the two feature maps extracted to double-branch structure are weighted and fused, realize surface solar irradiance prediction;The present application design scheme enhances the physical explainability of model, improves the precision, stability and generalization ability of solar irradiance short-term prediction.

Description

Technical Field

[0001] This invention relates to a solar irradiance prediction method that integrates physical priors and multi-scale features, belonging to the field of meteorological analysis and deep learning time series prediction technology for new energy power systems. Background Technology

[0002] In the global transition to low-carbon energy, photovoltaic (PV) power generation, as one of the most promising renewable energy sources, has seen its grid-connected capacity continuously increase. However, as the most crucial driver determining the power output of PV power plants, surface solar irradiance is subject to complex thermodynamic evolution in the atmosphere, aerosol scattering, and random drift of local cloud clusters, exhibiting extremely strong intermittent, fluctuating, and non-stationary characteristics. Drastic changes in solar irradiance across time and space directly trigger significant fluctuations in the output power of PV power plants, thus posing challenges to day-ahead and intraday dispatching, frequency regulation, and voltage stability control of the power system. Therefore, constructing high-precision ultra-short-term and short-term solar irradiance prediction models has become an indispensable core foundation for modern intelligent dispatching and energy management systems in power systems.

[0003] Solar irradiance prediction technology has attracted much attention from both academic and industrial communities both domestically and internationally. Early prediction methods mainly relied on persistent models or statistical techniques such as autoregressive moving averages (ARIMA), but these methods have limited effectiveness in handling complex nonlinear meteorological fluctuations. With the development of artificial intelligence technology, sequence modeling based on deep learning has gradually become mainstream. For example, the Informer model utilizes a probabilistic sparse self-attention mechanism, providing an effective approach for processing historical meteorological big data. However, attention mechanism models, represented by Transformer and its variants, experience a quadratic increase in computational complexity with the sequence length when processing extremely long historical observation sequences. This leads to limitations in memory overhead, making it difficult to achieve real-time inference and deployment on edge devices with limited computing power.

[0004] In the field of multi-source meteorological data processing, the ERA5 global reanalysis dataset provides extremely rich physical dimension support for irradiance prediction. However, in actual engineering modeling, existing mainstream multi-dimensional time-series prediction models typically directly concatenate multi-source meteorological feature matrices, including sea level pressure, surface temperature, wind speed, and cloud cover at various layers, along the channel dimension before feeding them into the network for feature extraction. This non-decoupled end-to-end data processing approach ignores the essential differences in the evolution scales of different physical quantities in atmospheric dynamics. For example, air pressure and temperature, dominated by large-scale weather systems, are slowly evolving thermodynamic features, while low cloud cover and wind speed shear, dominated by local microphysical processes, are instantaneous, high-frequency fluctuating dynamic features. Coupled learning of features with vastly different evolutionary cycles within the same network space leads to mutual interference during backpropagation, weakening the model's physical interpretability and robustness under complex weather conditions.

[0005] To balance computational efficiency and long-term memory capability, Gu and Dao proposed the Mamba State-Space Model (SSM), which, through its selective scanning mechanism, successfully reduced the complexity of long sequence modeling to a linear level. The Mamba architecture demonstrates superior deployment potential compared to the Transformer. However, the measured surface solar irradiance signal is essentially a superposition of deterministic astronomical laws and random meteorological disturbances, including dramatic fluctuations caused by convective weather systems and a smoothed diurnal trend of Earth's rotation. Existing single-branch Mamba structures still face technical bottlenecks when processing such multi-scale composite signals: when attempting to fit a smooth macroscopic trend term, they often smooth out details of cloud disturbances with high-frequency characteristics, resulting in an "oversmoothed" prediction; conversely, they can easily lead to long-term trend deviations.

[0006] Furthermore, multi-step forecasting mechanisms generally lack the ability to adaptively perceive forecast lead time. In short-term forecasting tasks covering multiple time steps in the future (such as the next 1 to 24 hours), the dominant physical factors causing changes in irradiance are not static but undergo significant dynamic migration. For example, in the ultra-short-term period of 1 to 3 hours in the future, the movement and evolution of local cloud clusters are the dominant factors determining abrupt changes in irradiance; while in forecasts exceeding 12 hours in the future, the stable evolution of large-scale pressure and temperature fields dominates. Summary of the Invention

[0007] The technical problem to be solved by this invention is to provide a solar irradiance prediction method that integrates physical priors and multi-scale features. By adaptive temporal decomposition of the signal side on the data input side, combined with a state-space model with linear computational complexity and a step-size-aware dynamic fusion mechanism, this method solves the technical problems of multi-scale coupling of meteorological features, easy cancellation of long-term high and low frequency signals, and lack of multi-step prediction adaptability of the model. This improves the accuracy, stability and physical interpretability of short-term solar irradiance prediction.

[0008] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: The present invention designs a solar irradiance prediction method that integrates physical priors and multi-scale features, and executes the following steps A to F to obtain the surface solar irradiance prediction model corresponding to the target area, and executes step H to realize the prediction of surface solar irradiance of the target area.

[0009] Step A. Obtain the surface solar irradiance sequence and the preset meteorological data sequences for the target area corresponding to the preset historical time period. Align the surface solar irradiance and the meteorological data for each type along each moment in the preset historical time period, update the surface solar irradiance sequence and the meteorological data sequences for each type, and then proceed to Step B.

[0010] Step B. Divide the meteorological data sequences of each type into a subset of slowly varying meteorological quantity sequences that characterize large-scale stationary evolution. A subset of rapidly varying meteorological quantities that characterize local convection and transient disturbances It also obtains the solar zenith angle sequence for the target region corresponding to a preset historical time period and adds it to the slow-varying meteorological quantity sequence subset. Then proceed to step C;

[0011] Step C. According to the preset... The target sliding window of time length is used to calculate the trend components corresponding to the surface solar irradiance at each time in the historical period for the surface solar irradiance sequence, forming the surface solar irradiance trend sequence. Then, the detailed components corresponding to the surface solar irradiance at each time in the historical period are calculated to form the surface solar irradiance detailed sequence. Then, the process proceeds to step D.

[0012] Step D. By sliding the target sliding window across historical time periods, determine each analysis time period window, and determine the preset time interval between each analysis time period window and the future direction. And includes presets The prediction time window for the length of time, based on the trend sequence of surface solar irradiance, the detailed sequence of surface solar irradiance, and... Each sequence in Each sequence corresponds to the data at each time point in a single analysis time window. Combined with the data of the surface solar irradiance sequence at each time point in the corresponding prediction time window, a single sample is formed. Then, each sample is used to form a sample set, and then the process proceeds to step E.

[0013] Step E. Construct the network to be trained, which includes a trend modeling branch, a detail modeling branch, and a step-size-aware dynamic fusion decoding module, and then proceed to step F;

[0014] Step F. Based on each sample, using the data corresponding to the analysis time window in the sample as input and the data corresponding to the prediction time window in the sample as output, train the network to be trained to obtain the trained network, which is the surface solar irradiance prediction model corresponding to the target area.

[0015] Step H. Based on historical time periods, following steps A to D, obtain data for each moment within the target analysis time window corresponding to the target region in the historical time period. Process this data using a surface solar irradiance prediction model to obtain the time interval between the target region and the target analysis time window in the future direction. The predicted values ​​of surface solar irradiance at each time point within the target prediction window.

[0016] As a preferred technical solution of the present invention: in step A, the following processing and updating are performed on each meteorological type data sequence after data alignment and updating;

[0017] First, based on the preset threshold range corresponding to the weather type, and combined with statistics... In principle, outlier detection, cleaning, and updates are performed on meteorological data sequences.

[0018] Then, a time-dimension-based cubic spline interpolation method is applied to interpolate and update the meteorological data series.

[0019] Finally, the Z-Score normalization algorithm is used to normalize the meteorological data sequence, mapping it to a normalized data sequence with a mean of 0 and a unit variance, and updating the meteorological data sequence.

[0020] As a preferred technical solution of the present invention: in step B, for each moment in the preset historical time period, the following formula is used:

[0021] ;

[0022] Calculate the solar zenith angle of the target region at the corresponding time. This, in turn, constitutes a sequence of solar zenith angles for the target region corresponding to a preset historical period, wherein, Indicates the first historical period corresponding to the target region The zenith angle of the sun at that moment. Indicates the geographical latitude of the target region. Indicates the first historical period corresponding to the target region The solar declination angle at that moment. Indicates the first historical period corresponding to the target region The solar hour angle at a given moment.

[0023] As a preferred technical solution of the present invention: in step C, according to the preset... The target sliding window for the time-time sequence is first determined based on the Earth's surface solar irradiance sequence, using the following formula:

[0024] ;

[0025] The trend components of surface solar irradiance at each moment in the historical period were calculated. ,in, Indicates the first in a historical period The trend component of the surface solar irradiance at a given moment. This indicates the number of time points included in the target sliding window. Indicates the first in a historical period The Earth's surface solar irradiance at a given moment. Indicates rounding down;

[0026] Then follow the formula below:

[0027] ;

[0028] Calculate the detailed components corresponding to the surface solar irradiance at each moment in the historical period. ,in, Indicates the first in a historical period The solar irradiance on the Earth's surface at a given moment.

[0029] As a preferred embodiment of the present invention: in the network structure to be trained in step E, the input end of the trend modeling branch and the input end of the detail modeling branch constitute the two input ends of the network to be trained. The input end of the trend modeling branch is used to receive the surface solar irradiance trend component corresponding to each moment in the analysis time window, and the corresponding... The data for each sequence are processed by the trend modeling branch, which outputs a deep trend feature map; the input of the detail modeling branch receives the surface solar irradiance detail components corresponding to each time point in the analysis time window, as well as the corresponding... The data of each sequence are processed by the detail modeling branch and output deep detail feature maps; the output of the trend modeling branch and the output of the detail modeling branch are connected to the input of the step size-aware dynamic fusion decoding module, and the output of the step size-aware dynamic fusion decoding module constitutes the output of the network to be trained.

[0030] Both the trend modeling branch and the detail modeling branch include a channel stitching module and a feature extraction module with a built-in causal convolution module. The input of the channel stitching module constitutes the input of the modeling branch, the output of the channel stitching module is connected to the input of the feature extraction module, and the output of the feature extraction module constitutes the output of the modeling branch.

[0031] As a preferred embodiment of the present invention: the step-size-aware dynamic fusion decoding module includes a global average pooling layer, a sine / cosine position encoding layer, a feature interaction splicing layer, a multilayer perceptron (MLP), a normalization layer, a weighted calculation layer, and a fully connected decoding layer. The two inputs of the global average pooling layer constitute the first set of two inputs of the step-size-aware dynamic fusion decoding module, respectively connected to the outputs of the trend modeling branch and the detail modeling branch. The input of the sine / cosine position encoding layer is used to receive the interval between the analysis time window and the prediction time window. The outputs of the global average pooling layer and the sine / cosine position encoding layer are connected to the two inputs of the feature interaction splicing layer. The output of the feature interaction splicing layer is connected to the input of the multilayer perceptron (MLP). The output of the MLP is connected to the input of the normalization layer. The output of the normalization layer is connected to one of the inputs of the weighted calculation layer. The normalization layer outputs weights corresponding to the depth trend feature map and the depth detail feature map to the weighted calculation layer. The outputs of the trend modeling branch and the detail modeling branch are simultaneously connected to the other input of the weighted calculation layer. The weighted calculation layer performs weighted processing on the received depth trend feature map and depth detail feature map according to the received weights to obtain the fused reconstructed feature map. The output of the weighted calculation layer is connected to the input of the fully connected decoding layer. The output of the fully connected decoding layer constitutes the output of the step-aware dynamic fusion decoding module.

[0032] As a preferred technical solution of the present invention: the preset meteorological data of various types include downward surface solar radiation, direct surface solar radiation, 2-meter temperature, surface temperature, 2-meter dew point temperature, mean sea level air pressure, surface air pressure, total column water vapor, boundary layer height, total precipitation, total cloud cover, low cloud cover, middle cloud cover, high cloud cover, cloud base height, 10-meter wind speed U component, and 10-meter wind speed V component.

[0033] In step B, the 2-meter temperature, surface temperature, 2-meter dew point temperature, mean sea level pressure, surface pressure, total column water vapor content, and boundary layer height are extracted into a subset of slowly varying meteorological quantity sequences. Total precipitation, total cloud cover, low cloud cover, middle cloud cover, high cloud cover, cloud base height, U-component of 10-meter wind speed, V-component of 10-meter wind speed, downward solar radiation at the surface, and direct solar radiation at the surface are classified into subsets of the rapidly varying meteorological quantity sequence. .

[0034] The solar irradiance prediction method integrating physical priors and multi-scale features described in this invention has the following technical advantages compared with existing technologies:

[0035] (1) The solar irradiance prediction method designed in this invention integrates physical priors and multi-scale features. First, it divides the meteorological data into fast-changing space and slow-changing space, and introduces the solar zenith angle as a deterministic benchmark. At the same time, it extracts the trend and detail terms of the surface solar irradiance sequence through adaptive signal decomposition. Then, it constructs a network to be trained. While maintaining linear computational complexity, it extracts deep trend feature maps and deep detail feature maps respectively using trend modeling branches and detail modeling branches. Then, it adds a step-size-aware dynamic fusion decoding module, and adaptively adjusts the weights of the two branches according to the prediction lead. It performs weighted fusion on the two feature maps extracted by the two-branch structure to realize the prediction of surface solar irradiance. The design scheme of this invention enhances the physical interpretability of the model and improves the accuracy, stability and generalization ability of short-term solar irradiance prediction.

[0036] (2) In the solar irradiance prediction method integrating physical priors and multi-scale features designed in this invention, the introduction of the solar zenith angle as a deterministic benchmark distinguishes between stationary evolution and instantaneous disturbance processes, which helps to reduce the difficulty of feature learning of heterogeneous data and enhance the physical interpretability of the model. The construction of trend modeling branch and detail modeling branch in the network to be trained, while maintaining linear computational complexity, effectively alleviates the feature "oversmoothing" phenomenon in multi-frequency signal extraction, takes into account both the macro-diurnal variation trend and the micro-fluctuation. Furthermore, the design of the step-size-aware dynamic fusion decoding module enables the model to adapt to the dynamic evolution of the influence of meteorological factors, adaptively adjust the fusion weights of fast and slow branches, improve the performance degradation of the traditional fixed weight mechanism in long-term prediction, and enhance the stability and generalization ability of multi-step prediction. Attached Figure Description

[0037] Figure 1 This is a schematic diagram of the overall architecture of the solar irradiance prediction method designed in this invention, which integrates physical priors and multi-scale features.

[0038] Figure 2 This is a schematic diagram of the architecture of the network to be trained in the design of this invention. Detailed Implementation

[0039] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0040] The solar irradiance prediction method designed in this invention integrates physical priors and multi-scale features. In practical applications, such as... Figure 1 As shown, the specific design involves performing steps A to F to obtain the surface solar irradiance prediction model for the target area.

[0041] Step A. Obtain the surface solar irradiance sequence and the preset meteorological data sequences for the target area corresponding to the preset historical time period. Align the surface solar irradiance and the meteorological data for each type along each moment in the preset historical time period. Update the surface solar irradiance sequence and the meteorological data sequences for each type. Then, perform the following processing update for each meteorological data sequence after data alignment and update. Proceed to Step B.

[0042] First, based on the preset threshold range corresponding to the weather type, and combined with statistics... In principle, outlier detection, cleaning, and updates are performed on meteorological data sequences.

[0043] Then, a time-dimension-based cubic spline interpolation method is applied to interpolate and update the meteorological data series.

[0044] Finally, the Z-Score normalization algorithm is used to normalize the meteorological data sequence, mapping it to a normalized data sequence with a mean of 0 and a unit variance, and updating the meteorological data sequence.

[0045] In practice, when performing standardization, for the first... Meteorological data sequences of various types Its standardized mapping formula is as follows:

[0046] ;

[0047] in, express The corresponding statistical expected mean, express The corresponding statistical standard deviation To prevent extremely small positive numbers with a denominator of zero.

[0048] Regarding the preset meteorological data of various types, the actual application design adopts 17 core variables of ERA5 multi-source meteorological parameters, covering multiple physical dimensions such as radiation, heat, dynamics and water vapor. Specifically, these include: downward surface solar radiation, direct surface solar radiation, 2-meter temperature, surface temperature, 2-meter dew point temperature, mean sea level pressure, surface pressure, total column water vapor, boundary layer height, total precipitation, total cloud cover, low cloud cover, middle cloud cover, high cloud cover, cloud base height, 10-meter wind speed U component, and 10-meter wind speed V component.

[0049] Step B. Divide the meteorological data sequences of each type into a subset of slowly varying meteorological quantity sequences that characterize large-scale stationary evolution. A subset of rapidly varying meteorological quantities that characterize local convection and transient disturbances In practical applications, this involves performing a priori decoupling based on the timescale of atmospheric physical dynamics evolution. Utilizing the evolution time constant property in the atmospheric thermodynamic equations, a subset of slowly varying meteorological quantities—including 2-meter temperature, surface temperature, 2-meter dew point temperature, mean sea level pressure, surface pressure, total column water vapor content, and boundary layer height—is extracted and decomposed. Total precipitation, total cloud cover, low cloud cover, middle cloud cover, high cloud cover, cloud base height, U-component of 10-meter wind speed, V-component of 10-meter wind speed, downward solar radiation at the surface, and direct solar radiation at the surface are classified into subsets of the rapidly varying meteorological quantity sequence. .

[0050] Then, for each moment within the preset historical time period, the following formula is applied:

[0051] ;

[0052] Calculate the solar zenith angle of the target region at the corresponding time. This leads to the formation of a solar zenith angle sequence for the target region corresponding to a preset historical period, which is then added to the subset of slowly varying meteorological quantities. In the middle, the reference envelope of irradiance under clear sky conditions undisturbed by weather is used as an anchor; among which, Indicates the first historical period corresponding to the target region The zenith angle of the sun at that moment. Indicates the geographical latitude of the target region. Indicates the first historical period corresponding to the target region The solar declination angle at that moment. Indicates the first historical period corresponding to the target region The solar hour angle at each moment is then determined, and then step C is performed.

[0053] Step C. According to the preset... The target sliding window for the length of time, such as a target sliding window containing 24 whole hours, i.e., a 24-hour target sliding window, is first calculated for the Earth's surface solar irradiance sequence using the following formula:

[0054] ;

[0055] The trend components of surface solar irradiance at each moment in the historical period were calculated. This constitutes a trend sequence of Earth's surface solar irradiance, among which... Indicates the first in a historical period The trend component of the surface solar irradiance at a given moment. This indicates the number of time points included in the target sliding window. Indicates the first in a historical period The Earth's surface solar irradiance at a given moment. This indicates rounding down to the nearest integer.

[0056] Then follow the formula below:

[0057] ;

[0058] Calculate the detailed components corresponding to the surface solar irradiance at each moment in the historical period. This constitutes a detailed sequence of solar irradiance on the Earth's surface, among which... Indicates the first in a historical period The surface solar irradiance at each moment is then recorded, and then step D is performed.

[0059] Step D. By sliding the target sliding window across historical time periods, determine each analysis time period window, and determine the preset time interval between each analysis time period window and the future direction. And includes presets The prediction time window for the length of time, based on the trend sequence of surface solar irradiance, the detailed sequence of surface solar irradiance, and... Each sequence in Each sequence corresponds to the data at each time point in a single analysis time window. Combined with the data of the surface solar irradiance sequence at each time point in the corresponding prediction time window, a single sample is formed. Then, each sample is used to form a sample set, and then the process proceeds to step E.

[0060] Step E. Construct a training network that includes a trend modeling branch, a detail modeling branch, and a step-size-aware dynamic fusion decoding module, and then proceed to step F.

[0061] In practical applications, specific designs are made for the network structure to be trained, such as... Figure 2 As shown, the input terminals of the trend modeling branch and the detail modeling branch constitute the two input terminals of the network to be trained. The input terminal of the trend modeling branch is used to receive the trend component of surface solar irradiance corresponding to each moment in the analysis time window. and corresponding The data for each sequence are processed by the trend modeling branch, which outputs a deep trend feature map. The input to the detailed modeling branch receives the detailed components of surface solar irradiance at each time point within the analysis time window. and corresponding The data from each sequence are processed by the detail modeling branch, which outputs a deep detail feature map. The outputs of the trend modeling branch and the detail modeling branch are connected to the input of the step-size-aware dynamic fusion decoding module, and the output of the step-size-aware dynamic fusion decoding module constitutes the output of the network to be trained.

[0062] Both the trend modeling branch and the detail modeling branch include a channel concatenation module and a feature extraction module with a built-in causal convolution module. The input of the channel concatenation module forms the input of the modeling branch, and the output of the channel concatenation module is connected to the input of the feature extraction module. The output of the feature extraction module forms the output of the modeling branch. In practical applications, each modeling branch uses a state-space model (Mamba) with linear computational complexity and a selective scanning mechanism. and After piecing together the data, we can further study the macroscopic evolutionary patterns and... and After stitching, the characteristics of sudden nonlinear meteorological disturbances are further captured.

[0063] The step-size-aware dynamic fusion decoding module is specifically designed with a global average pooling layer, a sine / cosine position encoding layer, a feature interaction concatenation layer, a multilayer perceptron (MLP), a normalization layer, a weighted calculation layer, and a fully connected decoding layer. The two inputs of the global average pooling layer constitute the first set of two inputs to the step-size-aware dynamic fusion decoding module, respectively connecting to the outputs of the trend modeling branch and the detail modeling branch. The input of the sine / cosine position encoding layer is used to receive the interval between the analysis window and the prediction window. The sine and cosine position coding layers map the vectors to high-dimensional conditionally aware embedding vectors. The outputs of the global average pooling layer and the sine / cosine position encoding layer are connected to the two inputs of the feature interaction concatenation layer. The output of the feature interaction concatenation layer is connected to the input of the multilayer perceptron (MLP). The output of the MLP is connected to the input of the normalization layer. The output of the normalization layer is connected to one of the inputs of the weighted calculation layer. The normalization layer outputs weights corresponding to the depth trend feature map and the depth detail feature map to the weighted calculation layer, respectively, as shown in the following formula:

[0064] ;

[0065] Obtaining deep trend feature maps Corresponding weights The outputs of both the trend modeling branch and the detail modeling branch are connected to the other input of the weighted calculation layer. The weighted calculation layer then performs weighted processing on the received deep trend feature map and deep detail feature map according to the received weights to obtain the fused reconstructed feature map. That is, the following formula:

[0066] ;

[0067] The output of the weighted calculation layer is connected to the input of the fully connected decoding layer, and the output of the fully connected decoding layer constitutes the output of the step-aware dynamic fusion decoding module.

[0068] The aforementioned design mechanism causes the model to bias weights towards capturing transient features when the prediction step size is short. As the prediction step size increases, the model weights shift towards representing the baseline trend characteristics on a macroscopic scale. The system is tilted to adapt to the dynamic migration patterns of the influence of meteorological factors, ultimately achieving stability in multi-step forecasting.

[0069] Step F. Based on each sample, using the data corresponding to the analysis time window in the sample as input and the data corresponding to the prediction time window in the sample as output, train the network to be trained. The mean square error (MSE) between the actual surface solar irradiance and the predicted surface solar irradiance is used as the loss function. The optimization algorithm (AdamW) is used in combination with the early stopping mechanism to complete the end-to-end iterative update of the network parameters, and the trained network is obtained, which is the surface solar irradiance prediction model corresponding to the target area.

[0070] Following steps A to F above, a prediction model for the surface solar irradiance of the target area is obtained. In practical applications, step H is then performed to predict the surface solar irradiance of the target area.

[0071] Step H. Based on historical time periods, following steps A to D, obtain data for each moment within the target analysis time window corresponding to the target region in the historical time period. Process this data using a surface solar irradiance prediction model to obtain the time interval between the target region and the target analysis time window in the future direction. The predicted values ​​of surface solar irradiance at each time point within the target prediction window.

[0072] The solar irradiance prediction method designed in this invention integrates physical priors and multi-scale features. First, it divides meteorological data into fast-varying and slow-varying spaces, introducing the solar zenith angle as a deterministic benchmark. Simultaneously, it extracts the trend and detail terms of the surface solar irradiance sequence through adaptive signal decomposition. Then, it constructs a training network, initially using trend modeling and detail modeling branches to maintain linear computational complexity while extracting deep trend feature maps and deep detail feature maps respectively. Next, it adds a step-size-aware dynamic fusion decoding module, adaptively adjusting the weights of the two branches based on the prediction lead. The two feature maps extracted from the two-branch structure are then weighted and fused to achieve surface solar irradiance prediction. This invention enhances the physical interpretability of the model and improves the accuracy, stability, and generalization ability of short-term solar irradiance prediction.

[0073] In the design scheme, the introduction of the solar zenith angle as a deterministic benchmark distinguishes between stationary evolution and instantaneous disturbance processes, which helps reduce the difficulty of feature learning for heterogeneous data and enhances the physical interpretability of the model. The construction of trend modeling branches and detail modeling branches in the network to be trained effectively alleviates the "oversmoothing" phenomenon in multi-frequency signal extraction while maintaining linear computational complexity, taking into account both the macroscopic daily variation trend and the microscopic fluctuation. Furthermore, the design of a step-size-aware dynamic fusion decoding module enables the model to adapt to the dynamic evolution of the influence of meteorological factors and adaptively adjust the fusion weights of fast and slow branches, improving the performance degradation of the traditional fixed weight mechanism in long-term forecasting and enhancing the stability and generalization ability of multi-step forecasting.

[0074] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.

Claims

1. A method for predicting solar irradiance that integrates physical priors and multi-scale features, characterized in that: Perform steps A to F to obtain the surface solar irradiance prediction model for the target area, and then perform step H to predict the surface solar irradiance of the target area. Step A. Obtain the surface solar irradiance sequence and the preset meteorological data sequences for the target area corresponding to the preset historical time period. Align the surface solar irradiance and the meteorological data for each type along each moment in the preset historical time period, update the surface solar irradiance sequence and the meteorological data sequences for each type, and then proceed to Step B. Step B. Divide the meteorological data sequences of each type into a subset of slowly varying meteorological quantity sequences that characterize large-scale stationary evolution. A subset of rapidly varying meteorological quantities that characterize local convection and transient disturbances It also obtains the solar zenith angle sequence for the target region corresponding to a preset historical time period and adds it to the slow-varying meteorological quantity sequence subset. Then proceed to step C; Step C. According to the preset... The target sliding window of time length is used to calculate the trend components corresponding to the surface solar irradiance at each time in the historical period for the surface solar irradiance sequence, forming the surface solar irradiance trend sequence. Then, the detailed components corresponding to the surface solar irradiance at each time in the historical period are calculated to form the surface solar irradiance detailed sequence. Then, the process proceeds to step D. Step D. By sliding the target sliding window across historical time periods, determine each analysis time period window, and determine the preset time interval between each analysis time period window and the future direction. And includes presets The prediction time window for the length of time, based on the trend sequence of surface solar irradiance, the detailed sequence of surface solar irradiance, and... Each sequence in Each sequence corresponds to the data at each time point in a single analysis time window. Combined with the data of the surface solar irradiance sequence at each time point in the corresponding prediction time window, a single sample is formed. Then, each sample is used to form a sample set, and then the process proceeds to step E. Step E. Construct the network to be trained, which includes a trend modeling branch, a detail modeling branch, and a step-size-aware dynamic fusion decoding module, and then proceed to step F; Step F. Based on each sample, using the data corresponding to the analysis time window in the sample as input and the data corresponding to the prediction time window in the sample as output, train the network to be trained to obtain the trained network, which is the surface solar irradiance prediction model corresponding to the target area. Step H. Based on historical time periods, following steps A to D, obtain data for each moment within the target analysis time window corresponding to the target region in the historical time period. Process this data using a surface solar irradiance prediction model to obtain the time interval between the target region and the target analysis time window in the future direction. The predicted values ​​of surface solar irradiance at each time point within the target prediction window.

2. The solar irradiance prediction method integrating physical priors and multi-scale features according to claim 1, characterized in that: In step A, the following processing and updates are performed on each meteorological data sequence after data alignment and updates; First, based on the preset threshold range corresponding to the weather type, and combined with statistics... In principle, outlier detection, cleaning, and updates are performed on meteorological data sequences. Then, a time-dimension-based cubic spline interpolation method is applied to interpolate and update the meteorological data series. Finally, the Z-Score normalization algorithm is used to normalize the meteorological data sequence, mapping it to a normalized data sequence with a mean of 0 and a unit variance, and updating the meteorological data sequence.

3. The solar irradiance prediction method integrating physical priors and multi-scale features according to claim 1, characterized in that: In step B, for each moment within the preset historical time period, the following formula is used: ； Calculate the solar zenith angle of the target region at the corresponding time. This, in turn, constitutes a sequence of solar zenith angles for the target region corresponding to a preset historical period, wherein, Indicates the first historical period corresponding to the target region The zenith angle of the sun at that moment. Indicates the geographical latitude of the target region. Indicates the first historical period corresponding to the target region The solar declination angle at that moment. Indicates the first historical period corresponding to the target region The solar hour angle at a given moment.

4. The solar irradiance prediction method integrating physical priors and multi-scale features according to claim 1, characterized in that: In step C, according to the preset... The target sliding window for the time-time sequence is first determined based on the Earth's surface solar irradiance sequence, using the following formula: ； The trend components of surface solar irradiance at each moment in the historical period were calculated. ,in, Indicates the first in a historical period The trend component of the surface solar irradiance at a given moment. This indicates the number of time points included in the target sliding window. Indicates the first in a historical period The Earth's surface solar irradiance at a given moment. Indicates rounding down; Then follow the formula below: ； Calculate the detailed components corresponding to the surface solar irradiance at each moment in the historical period. ,in, Indicates the first in a historical period The solar irradiance on the Earth's surface at a given moment.

5. The solar irradiance prediction method integrating physical priors and multi-scale features according to claim 1, characterized in that: In step E, the input of the trend modeling branch and the input of the detail modeling branch constitute the two inputs of the network to be trained. The input of the trend modeling branch is used to receive the trend component of the surface solar irradiance corresponding to each moment in the analysis time window, and the corresponding... The data for each sequence are processed by the trend modeling branch, which outputs a deep trend feature map. The input to the detailed modeling branch receives the detailed components of the surface solar irradiance at each time point within the analysis time window, as well as the corresponding... The data of each sequence are processed by the detail modeling branch and output deep detail feature maps; the output of the trend modeling branch and the output of the detail modeling branch are connected to the input of the step size-aware dynamic fusion decoding module, and the output of the step size-aware dynamic fusion decoding module constitutes the output of the network to be trained. Both the trend modeling branch and the detail modeling branch include a channel stitching module and a feature extraction module with a built-in causal convolution module. The input of the channel stitching module constitutes the input of the modeling branch, the output of the channel stitching module is connected to the input of the feature extraction module, and the output of the feature extraction module constitutes the output of the modeling branch.

6. The solar irradiance prediction method integrating physical priors and multi-scale features according to claim 1, characterized in that: The step-size-aware dynamic fusion decoding module includes a global average pooling layer, a sine / cosine position encoding layer, a feature interaction splicing layer, a multilayer perceptron (MLP), a normalization layer, a weighted calculation layer, and a fully connected decoding layer. The two inputs of the global average pooling layer constitute the first set of two inputs of the step-size-aware dynamic fusion decoding module, respectively connecting to the outputs of the trend modeling branch and the detail modeling branch. The input of the sine / cosine position encoding layer is used to receive the interval between the analysis time window and the prediction time window. The outputs of the global average pooling layer and the sine / cosine position encoding layer are connected to the two inputs of the feature interaction splicing layer. The output of the feature interaction splicing layer is connected to the input of the multilayer perceptron (MLP). The output of the MLP is connected to the input of the normalization layer. The output of the normalization layer is connected to one of the inputs of the weighted calculation layer. The normalization layer outputs weights corresponding to the depth trend feature map and the depth detail feature map to the weighted calculation layer. The outputs of the trend modeling branch and the detail modeling branch are simultaneously connected to the other input of the weighted calculation layer. The weighted calculation layer performs weighted processing on the received depth trend feature map and depth detail feature map according to the received weights to obtain the fused reconstructed feature map. The output of the weighted calculation layer is connected to the input of the fully connected decoding layer. The output of the fully connected decoding layer constitutes the output of the step-aware dynamic fusion decoding module.

7. The solar irradiance prediction method integrating physical priors and multi-scale features according to claim 1, characterized in that: The preset meteorological data include downward surface solar radiation, direct surface solar radiation, 2-meter temperature, surface temperature, 2-meter dew point temperature, mean sea level pressure, surface pressure, total column water vapor, boundary layer height, total precipitation, total cloud cover, low cloud cover, middle cloud cover, high cloud cover, cloud base height, 10-meter wind speed U component, and 10-meter wind speed V component. In step B, the 2-meter temperature, surface temperature, 2-meter dew point temperature, mean sea level pressure, surface pressure, total column water vapor content, and boundary layer height are extracted into a subset of slowly varying meteorological quantity sequences. Total precipitation, total cloud cover, low cloud cover, middle cloud cover, high cloud cover, cloud base height, U-component of 10-meter wind speed, V-component of 10-meter wind speed, downward solar radiation at the surface, and direct solar radiation at the surface are classified into subsets of the rapidly varying meteorological quantity sequence. .

Images