Time-step-by-time-step total solar radiation multi-time-domain prediction method, product and equipment

Abstract

This invention provides a method, product, and device for multi-temporal-domain prediction of total solar radiation at time steps. The method includes: acquiring and preprocessing meteorological observation data from multiple data sources to obtain model feature data; constructing multiple sliding window sequences based on the model feature data and multiple preset target prediction time domains; training multiple structurally distinct deep learning base models using the sliding window sequences; feeding the basic prediction sequences output by the multiple deep learning base models into an attention stacking fusion layer for fusion processing to obtain a fused prediction sequence; and outputting multi-step solar radiation prediction results across multiple time domains based on the fused prediction sequence. This method enables the independent learning of optimal base model combination weights for each prediction step in multi-temporal-domain prediction of total solar radiation, adaptively capturing the advantages of each model at different prediction distances, and achieving better prediction accuracy across all time domains.

Description

Technical Field

[0001] This invention relates to the field of total solar radiation prediction technology, and in particular to a time-step multi-time-domain prediction method, product, and equipment for total solar radiation. Background Technology

[0002] Total solar radiation is the primary source of energy on Earth's surface and a core driving force for solar energy utilization systems such as photovoltaic power generation and solar thermal utilization. With the rapid development of the global new energy industry, accurate prediction of total solar radiation is of paramount importance for the optimized scheduling of photovoltaic power plants, the rational allocation of energy storage systems, electricity market trading decisions, and the safe and stable operation of the power grid.

[0003] However, total solar radiation is influenced by a variety of complex factors, including cloud movement, aerosol distribution, water vapor content, and weather system evolution, exhibiting significant randomness and making accurate prediction challenging. Especially under complex terrain and variable weather conditions, short-term fluctuations in total solar radiation are more pronounced, and traditional statistical regression and purely physical models show significant errors under such conditions. Therefore, improving the accuracy and reliability of total solar radiation prediction to meet application needs across different time scales is a crucial technical problem that urgently needs to be solved in this field. Summary of the Invention

[0004] In view of the above problems, the present invention proposes a method, product and device for multi-temporal-domain prediction of total solar radiation at time steps to overcome the above problems or at least partially solve the above problems.

[0005] One objective of this invention is to enable each prediction step in multi-step total solar radiation prediction to adaptively obtain a set of optimal combined weights for each base model, instead of using a fixed single weight, thereby improving the accuracy of multi-temporal prediction of total solar radiation.

[0006] Specifically, this invention provides a multi-temporal-domain prediction method for total solar radiation step-by-step, comprising: Meteorological observation data from multiple data sources are acquired and preprocessed to obtain model feature data; Multiple sliding window sequences are constructed based on model feature data and preset multiple target prediction time domains; Multiple deep learning base models with distinct structures are trained using sliding window sequences. These deep learning base models are used to independently output a basic prediction sequence containing multiple prediction steps. The basic prediction sequences output by multiple deep learning base models are fed into the attention stack fusion layer for fusion processing to obtain a fused prediction sequence. The attention stack fusion layer maintains a set of weights for each deep learning base model for each prediction step in the basic prediction sequence, and is used to receive the basic prediction sequences output by each deep learning base model. After normalizing the weights of each prediction step, the predicted values ​​of each deep learning base model are weighted and summed according to the prediction step. The multi-step total solar radiation prediction results in multiple time domains are output based on the fused prediction sequence.

[0007] Optionally, the steps of feeding the basic prediction sequences output by multiple deep learning base models into an attention stacking fusion layer for fusion processing to obtain the fused prediction sequence include: Construct a weight parameter matrix. The number of rows in the weight parameter matrix is ​​equal to the number of prediction steps, and the number of columns is equal to the number of deep learning base models. The elements in the weight parameter matrix represent the unnormalized logarithm of the deep learning base models at the corresponding prediction steps. The weight parameter matrix is ​​initialized to zero. The weight parameters corresponding to each prediction step are gradually normalized through a normalized exponential function to obtain the combination coefficients corresponding to the prediction step. The combination coefficients are non-negative and sum to 1. The combination coefficients represent the contribution ratio of each deep learning base model in the prediction step. Based on the combination coefficients corresponding to each prediction step, the predicted values ​​of each deep learning base model at the corresponding prediction step are weighted and summed step by step to obtain the fused prediction sequence.

[0008] Optionally, the data source includes: satellite remote sensing data, ground observation station data, and observation geometry data; Satellite remote sensing data includes: multi-channel shortwave albedo and multi-channel infrared brightness temperature; Ground observation station data includes: air temperature, wind direction, wind speed, relative humidity, air pressure, surface temperature, visibility, and historical total solar radiation from multiple observation stations; The observed geometric angle data include: solar zenith angle, solar azimuth angle, satellite zenith angle, and satellite azimuth angle; The sliding window sequence includes the input window sequence, the prediction window sequence, and the sliding window step size; Deep learning base models include convolutional neural network-bidirectional long short-term memory network-attention cascade model, self-attention encoder model, and bidirectional long short-term memory network-additive attention model.

[0009] Optionally, the step of constructing multiple sets of sliding window sequences based on model feature data and multiple preset target prediction time domains includes: Based on multiple preset target prediction time domains, determine the input window length, prediction window length, and sliding window step size corresponding to each target prediction time domain; Based on the input window length, prediction window length, and sliding window step size, the input window sequence and the corresponding prediction window sequence are extracted from the model feature data, and the input window length is set to cover at least two complete daily cycles.

[0010] Optionally, before the step of feeding the basic prediction sequences output by multiple deep learning base models into the attention stack fusion layer for fusion processing, the following steps are also included: The meteorological observation data is divided into a training set and a validation set, with each set containing different stations. The training set is used to train the deep learning base model, while the validation set is used to evaluate the model's generalization ability.

[0011] Optionally, the training steps for the attention stacking fusion layer include: The predicted sequences output by each deep learning base model on the validation set are obtained as training inputs, and the real observations are used as supervision signals. Determine the loss function and optimizer to be used for training; The weight parameters of the attention stacked fusion layer are updated based on the training input and supervision signals. By dividing the internal validation set from the validation set predictions and using an early stopping strategy, overfitting of the attention stacking fusion layer can be prevented.

[0012] Optionally, after the step of feeding the basic prediction sequences output by multiple deep learning base models into the attention stacking fusion layer for fusion processing, the method further includes: The prediction performance indices of the fused prediction sequence and each basic prediction sequence are calculated. The prediction performance indices include global indices, daytime indices, and sub-indices. The global indices are used to reflect the overall error of the entire prediction sequence. The daytime indices are used to remove the influence of zero-value samples at night and reflect the daytime prediction performance. The sub-indices are statistically analyzed according to the validation site, season, and prediction step, and are used to locate the prediction performance under different climatic conditions and different prediction distances.

[0013] Optionally, the preprocessing steps include: Quality control is performed on meteorological observation data to eliminate abnormal data; Fill in missing values ​​in meteorological observation data; Meteorological observation data are merged based on timestamps to form a unified feature table; The dynamic features, static features, and predicted labels in the feature table are standardized to obtain the model feature data.

[0014] According to another aspect of the invention, a computer program product is also provided, comprising a computer program that, when executed by a processor, implements the steps of any of the above-described time-step multi-temporal prediction methods for total solar radiation.

[0015] According to another aspect of the present invention, a computer device is also provided, including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the computer program to implement the steps of any of the above-described time-step multi-temporal prediction methods for total solar radiation.

[0016] The present invention provides a time-step multi-temporal prediction method for total solar radiation. First, meteorological observation data from multiple data sources are acquired and preprocessed to obtain model feature data. Then, multiple sliding window sequences are constructed based on the model feature data and multiple preset target prediction time domains. Next, multiple structurally distinct deep learning base models are trained using these sliding window sequences. Each deep learning base model independently outputs a basic prediction sequence containing multiple prediction steps. The basic prediction sequences output by the multiple deep learning base models are then fed into an attention stacking fusion layer for fusion processing to obtain a fused prediction sequence. The attention stacking fusion layer independently maintains a set of weights for each deep learning base model for each prediction step in the basic prediction sequence. It receives the basic prediction sequences output by each deep learning base model, normalizes the weights for each prediction step, and then performs a weighted summation of the predicted values ​​of each deep learning base model for each prediction step. Finally, the multi-step solar radiation prediction result in multiple time domains is output based on the fused prediction sequence. This method enables the independent learning of optimal base model combination weights for each prediction step in the multi-temporal prediction of total solar radiation, adaptively capturing the advantages of each model at different prediction distances, and achieving better prediction accuracy in all time domains.

[0017] The above and other objects, advantages and features of the present invention will become more apparent to those skilled in the art from the following detailed description of specific embodiments of the invention in conjunction with the accompanying drawings. Attached Figure Description

[0018] The following sections will describe some specific embodiments of the invention in detail by way of example and not limitation, with reference to the accompanying drawings. The same reference numerals in the drawings denote the same or similar parts or portions. Those skilled in the art should understand that these drawings are not necessarily drawn to scale. In the drawings: Figure 1 This is a flowchart illustrating a time-step multi-temporal-domain prediction method for total solar radiation according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the fusion process performed by the attention stacking fusion layer in the multi-temporal domain prediction method for time-step solar total radiation according to an embodiment of the present invention. Figure 3 This is a flowchart illustrating a multi-temporal-domain prediction method for time-step solar total radiation according to another embodiment of the present invention. Figure 4This is a schematic diagram of a computer program product according to an embodiment of the present invention; Figure 5 This is a schematic diagram of a computer-readable storage medium according to an embodiment of the present invention; and Figure 6 A schematic diagram of a computer device according to an embodiment of the present invention. Detailed Implementation

[0019] Those skilled in the art should understand that the embodiments described below are merely a part of the embodiments of the present invention, and not all of the embodiments of the present invention. These partial embodiments are intended to explain the technical principles of the present invention and are not intended to limit the scope of protection of the present invention. Based on the embodiments provided by the present invention, all other embodiments obtained by those skilled in the art without creative effort should still fall within the scope of protection of the present invention.

[0020] It should be noted that the logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be specifically implemented in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus or device (such as a computer-based system, a processor-included system or other system that can fetch and execute instructions from, an instruction execution system, apparatus or device).

[0021] The accuracy of total solar radiation prediction directly impacts the scheduling of photovoltaic power plants, energy storage configuration, and electricity market transactions. Variations in clouds, aerosols, and water vapor give radiation sequences strong randomness and high-frequency fluctuations, causing traditional statistical regression and purely physical models to have significant errors under variable weather conditions. In recent years, solutions to the problem of "how to improve the accuracy of multi-step prediction of total solar radiation" have at least the following drawbacks: (1) A single model has only one inductive bias. The model that is strongest in the short term is generally not the model that is strongest in the long term. It is difficult for it to be optimal in the entire prediction sequence. (2) Weighted average and grid search are types of fusion that use a set of weights for the entire prediction sequence, which cannot characterize the phenomenon that “the optimal combination of the near step and the far step is different”. (3) Selecting one model from multiple candidate models means losing the complementary information of other models and failing to obtain the integrated benefit of error cancellation. (4) If the meta-learner directly uses the predictions of the base model on the training set, it will be contaminated by the overfitting of the training set of the base model, resulting in mismatched fusion weights and degradation on the real data. (5) Most schemes only adjust parameters for a certain time domain. If the forecast duration is changed, the model must be redesigned. It is impossible to use a single framework to support the intraday scheduling, day-ahead planning and next-day forecast of photovoltaic power plants at the same time. (6) Randomly dividing the dataset will put data from the same site into training and validation at the same time. The model will "remember" the validation samples by taking advantage of the site inertia, and the accuracy will be overestimated. The true performance on sites that have not been seen before is unclear.

[0022] To address the aforementioned problems, this invention proposes a multi-temporal-domain prediction method for total solar radiation at each time step. Figure 1 This is a schematic diagram of the architecture of a time-step multi-temporal-domain prediction method for total solar radiation according to an embodiment of the present invention, as shown below. Figure 1 As shown, the process includes at least the following steps S101 to S105.

[0023] Step S101: Obtain meteorological observation data from multiple data sources and perform preprocessing to obtain model feature data. In some optional embodiments, this invention employs a multi-source data fusion strategy, mainly including three types of data sources: satellite remote sensing data, ground observation station data, and observation geometric angle data (generally including solar zenith angle, solar azimuth angle, satellite zenith angle, and satellite azimuth angle). The reason for using multi-source data is that a single data source is difficult to comprehensively characterize the various factors affecting total solar radiation: satellite data can provide large-scale cloud distribution and aerosol information, ground observation data can provide measured values ​​of local meteorological elements, and geometric angle data reflects the geometric characteristics of the total solar radiation transmission path. The three complement each other and can provide richer and more comprehensive input features for the prediction model.

[0024] Specifically, satellite data can be acquired from meteorological satellites (such as the Himawari-9 satellite), including multi-channel shortwave albedo and multi-channel infrared brightness temperature. The combination of shortwave albedo and infrared brightness temperature can effectively characterize the spatial distribution and dynamic changes of clouds. Ground-based observation station data includes meteorological elements such as air temperature, wind direction, wind speed, relative humidity, air pressure, surface temperature, and visibility, as well as historical total solar radiance data. Historical total solar radiance, as an autoregressive input, can provide the model with the inherent variation patterns of the radiation sequence. Observational geometry data is derived from the station's latitude and longitude, observation time, satellite orbit position, and sensor scanning geometry. This data determines the transmission path length of total solar radiation and directly affects radiation intensity. The solar zenith angle (SOZ) and solar azimuth angle (SOA) are derived from the station's latitude and longitude and observation time, while the satellite zenith angle (SAZ) and satellite azimuth angle (SAA) depend on the satellite orbit position and sensor scanning geometry.

[0025] Optionally, to ensure prediction accuracy, the aforementioned multi-source data can be preprocessed, specifically including the following steps: Quality control: Quality control is performed on meteorological observation data to remove outliers, such as observations exceeding physically reasonable ranges or outliers caused by instrument malfunctions. Quality control is fundamental to subsequent modeling, preventing outliers from interfering with model training.

[0026] Missing value imputation: Missing values ​​in meteorological observation data can be imputed. For fields with low missing values, forward or backward imputation can be used, while for fields still missing values, the column mean can be used as a fallback. Meteorological observation data often contains missing values ​​due to instrument maintenance, communication failures, etc. Appropriate missing value imputation can ensure the continuity of the data sequence and provide a complete data foundation for subsequent sliding window construction.

[0027] Feature merging: Meteorological observation data are merged according to timestamps to form a unified feature table. Since the sampling times of different data sources may differ, aligning by timestamps ensures that various features at the same moment correspond correctly.

[0028] Standardization: Dynamic features, static features, and predicted labels are standardized separately. Standardization can eliminate the differences in the units of measurement between different features, so that features of different units of measurement have equal contribution in model training, thereby accelerating model convergence and improving training stability.

[0029] Preprocessing provides high-quality input data for subsequent model training.

[0030] Step S102 involves constructing multiple sliding window sequences based on model feature data and preset multiple target prediction time domains. Each sliding window sequence includes an input window sequence, a prediction window sequence, and a sliding window step size. This invention supports multiple target prediction time domains, such as 4 hours, 12 hours, and 24 hours, with different input window lengths, prediction window lengths, and sliding window step size configurations corresponding to different prediction time domains. This design is based on the fact that a longer prediction time domain requires more historical context information, thus increasing the input window length accordingly; while the sliding window step size balances the number of samples and computational efficiency.

[0031] In some optional embodiments, the step of constructing multiple sets of sliding window sequences based on model feature data and multiple preset target prediction time domains generally includes: determining the input window length, prediction window length, and sliding window step size corresponding to each target prediction time domain based on the multiple preset target prediction time domains; extracting the input window sequence and the corresponding prediction window sequence from the model feature data based on the input window length, prediction window length, and sliding window step size, wherein the input window length is set to cover at least two complete daily cycles.

[0032] This invention sets the input window length to cover at least two complete diurnal cycles. This is because total solar radiation exhibits a significant diurnal cyclical variation; the model needs to observe at least two complete day-night cycles to learn periodic patterns such as sunrise, sunset, and noon peak, thereby more accurately predicting future radiation changes. By constructing a sliding window, continuous feature sequences can be transformed into input-output sample pairs that the model can directly process, providing a data foundation for subsequent base model training.

[0033] Step S103 involves training multiple structurally distinct deep learning base models using a sliding window sequence. These deep learning base models are used to independently output a basic prediction sequence containing multiple prediction steps. This invention employs multiple structurally distinct deep learning base models, the core idea of ​​which is to leverage the differences in inductive bias between different models to achieve complementary prediction results. Neural network models with different structures have different advantages and disadvantages when processing time-series data. By combining multiple models with different structures, their respective strengths can be utilized to compensate for the shortcomings of a single model, thereby improving the overall prediction accuracy.

[0034] In some optional embodiments, the deep learning base model of the present invention may include the following three types: The Convolutional Neural Network-Bilateral Long Short-Term Memory Network-Attention Cascade (CNN-BiLSTM-Attention, CBA) model consists of a dilated CNN (for multi-scale local pattern extraction), a BiLSTM (for long-range temporal modeling), and a multi-head self-attention (for adaptive aggregation) cascaded together, then fused with site embeddings and static features before being fed into the prediction head. Specifically, the dilated convolutions exponentially expand the receptive field by setting convolutional kernels with different dilation rates in parallel, maintaining temporal resolution without increasing parameters. This allows for the simultaneous capture of multi-scale local temporal patterns such as proximity mutations, intraday cycles, and cross-day trends. The bilateral long short-term memory network can model long-range temporal dependencies from both forward and backward directions, and the multi-head self-attention adaptively aggregates information from different time steps.

[0035] The Transformer self-attention encoder model is built upon a self-attention mechanism, employing a multi-layer, multi-head self-attention encoding structure. The advantage of self-attention lies in the fact that any two time steps in a long sequence can directly interact, unrestricted by distance, thus possessing a global perspective and maintaining good performance in long-term prediction. A Transformer is typically a 4-layer, 8-head self-attention encoder, where any two time steps in a long sequence can directly interact, excelling in its global perspective.

[0036] Bi-LSTM-Attention model: This model consists of a bi-LSTM network and Bahdanau additive attention. The model is relatively small, trains quickly, and has relatively stable error decay in the long time domain. It can serve as a useful supplement to other models and an ablation baseline.

[0037] There is no information exchange between the three base models; each is trained independently. This is done to ensure that each base model learns independent feature representations and prediction patterns, avoiding mutual interference that could lead to a decrease in diversity. This ensures that the complementary effects are fully utilized during fusion. By training multiple structurally distinct base models, multi-angle and multi-modal prediction results can be obtained, providing rich input for subsequent fusion layers and laying the foundation for improved accuracy.

[0038] Step S104 involves feeding the basic prediction sequences output by multiple deep learning base models into an attention stacking fusion layer for fusion processing to obtain a fused prediction sequence. The attention stacking fusion layer maintains an independent set of weights for each deep learning base model for each prediction step in the basic prediction sequence. It receives the basic prediction sequences output by each deep learning base model, normalizes the weights for each prediction step, and then performs a weighted summation of the predicted values ​​of each deep learning base model for each prediction step.

[0039] Traditional multi-model fusion methods typically assign a uniform set of weights to the entire prediction sequence, meaning that all prediction steps use the same combination ratio. However, in multi-step prediction tasks, using the same set of weights across the entire sequence limits further improvements in fusion accuracy.

[0040] The time-step attention stacking fusion mechanism proposed in this invention is precisely designed to solve this problem. Its core idea is to learn an optimal set of combined weights for each prediction step, allowing each time step to adaptively select the most suitable model combination.

[0041] Specifically, the steps of feeding the basic prediction sequences output by multiple deep learning base models into an attention stacking fusion layer for fusion processing to obtain the fused prediction sequence generally include: constructing a weight parameter matrix, where the number of rows in the weight parameter matrix equals the number of prediction steps and the number of columns equals the number of deep learning base models, and the elements in the weight parameter matrix represent the unnormalized logarithm of the deep learning base models at the corresponding prediction step, and the weight parameter matrix is ​​initialized to zero; gradually normalizing the weight parameters corresponding to each prediction step using a normalized exponential function to obtain the combination coefficients corresponding to the prediction step, where the combination coefficients are non-negative and sum to 1, and the combination coefficients represent the contribution ratio of each deep learning base model at the prediction step; and gradually weighting and summing the predicted values ​​of each deep learning base model at the corresponding prediction step according to the combination coefficients corresponding to each prediction step to obtain the fused prediction sequence.

[0042] In some optional embodiments, a weight parameter matrix is ​​constructed: a trainable weight parameter matrix of shape L×M is defined, where L is the number of prediction steps and M is the number of base models. Each element in the matrix represents the unnormalized logarithm of the weights of the corresponding base model at the corresponding prediction step. The weight matrix is ​​initialized to zero, corresponding to the initial state of equal-weighted fusion of all base models. The advantages of this design are: the number of parameters is extremely small, only L×M; for example, for a 4-hour prediction, only 24×3=72 parameters are needed; the small number of parameters means that the fusion layer will hardly overfit, and it can be trained stably even with limited validation data.

[0043] Stepwise normalization: The weight parameters corresponding to each prediction step are normalized using a normalized exponential function to obtain the combination coefficients for that prediction step. Normalization using the normalized exponential function ensures that the combination coefficients for each prediction step are non-negative and sum to 1, possessing a clear probabilistic meaning and can be directly interpreted as the contribution ratio of each base model to that prediction step. This inherent interpretability is a significant advantage of this invention. After training, simply printing the weight matrix provides a clear view of which model is primarily responsible for which step and whether the long-term steps are balanced across multiple models, facilitating attribution analysis of anomalies and model debugging.

[0044] Stepwise weighted summation: Based on the combination coefficients corresponding to each prediction step, the predicted values ​​of each base model at that prediction step are weighted and summed to obtain the fused prediction sequence. Each prediction step only uses the weights of its corresponding row, and the prediction steps do not affect each other.

[0045] The training steps for an attention-stacked fusion layer typically include: obtaining the predicted sequences output by each deep learning base model on the validation set as training input, and using real observations as supervision signals; determining the loss function and optimizer used for training; updating the weight parameters of the attention-stacked fusion layer based on the training input and supervision signals; and partitioning an internal validation set from the validation set predictions and employing an early stopping strategy to prevent overfitting of the attention-stacked fusion layer. This is because base models often overfit on the training set, especially long short-term memory models with a large number of parameters, where the error on the training set is often much smaller than the actual generalization error. If the fusion layer is trained directly using the training set predictions, it will overestimate the weights of the severely overfitted model, leading to a degradation in fusion performance on unseen real data. Using validation set predictions to train the fusion layer, however, better reflects the true generalization ability of the base models since the validation set data was not used in the training, making the weights learned by the fusion layer more reliable and its performance on real data more stable. This strategy effectively avoids the contamination of the fusion weights by base model overfitting.

[0046] During training of the fusion layer, the Huber loss function can be used to calculate the prediction error, and an adaptive moment estimation optimizer can be used to iteratively update the weight parameters using gradient descent. The Huber loss is insensitive to outliers and is more robust than mean squared error. Furthermore, an internal validation set can be further partitioned from the validation set predictions. This can be used in conjunction with an early stopping strategy, stopping training early when the loss on the internal validation set stops decreasing for several consecutive rounds, thus preventing the fusion layer from overfitting on the validation set.

[0047] By using a time-step attention stacking fusion mechanism, each prediction step can adaptively obtain the optimal model combination, giving full play to the advantages of different base models at different prediction distances, thereby significantly improving the overall prediction accuracy.

[0048] In some alternative embodiments, before the step of feeding the basic prediction sequences output by multiple deep learning base models into the attention stack fusion layer for fusion processing, the method may further include: dividing multiple observation stations corresponding to the meteorological observation data into a training set and a validation set, wherein the stations in the training set and the validation set are different from each other. The training set is used to train the deep learning base models, and the validation set is used to evaluate the generalization ability of the models. This station-based spatial partitioning method ensures that the data of the validation stations are not involved in the training at all, thereby truly evaluating the model's spatial generalization ability to unseen stations and avoiding the problem of overestimating accuracy due to station inertia caused by random partitioning.

[0049] Optionally, after the step of feeding the basic prediction sequences output by multiple deep learning base models into the attention stacking fusion layer for fusion processing, the following may also be included: calculating the prediction performance indicators of the fused prediction sequence and each basic prediction sequence. The prediction performance indicators include global indicators, daytime indicators, and sub-indicators. The global indicators are used to reflect the overall error of the entire prediction sequence. The daytime indicators are used to remove the influence of zero-value samples at night and reflect the daytime prediction performance. The sub-indicators are statistically analyzed according to the validation site, season, and prediction step, and are used to locate the prediction performance under different climatic conditions and different prediction distances.

[0050] Step S105: Output the multi-step total solar radiation prediction results in multiple time domains based on the fused prediction sequence.

[0051] This invention employs a unified parametric framework to support multiple prediction time domains. By configuring two parameters—input window length and prediction window length—it is possible to switch between different prediction durations without altering the main model structure. The significance of this design is that a single methodology can simultaneously support various business needs of photovoltaic power plants, such as intraday scheduling, day-ahead planning, and next-day forecasting, eliminating the repetitive work of maintaining a separate model for each time domain and significantly reducing development and maintenance costs.

[0052] In summary, this invention effectively improves the accuracy and generalization ability of multi-step prediction of total solar radiation through techniques such as a time-step attention stacking fusion mechanism, a validation set training strategy, and a unified multi-temporal framework, while also possessing good interpretability and engineering practicality. Experimental results show that the fusion model of this invention outperforms the best single model in all temporal domains, and the longer the prediction time domain, the more significant the fusion gain. On validation sites covering multiple climate zones and different altitudes, the fusion model achieves accuracy improvements at the vast majority of sites, demonstrating good spatial generalization ability.

[0053] Figure 2 This is a schematic diagram illustrating the fusion process of the attention stacking fusion layer in a time-step multi-temporal prediction method for total solar radiation according to an embodiment of the present invention. Figure 2 As shown, the process includes at least the following steps S201 to S203.

[0054] Step S201: Construct the weight parameter matrix.

[0055] In some optional embodiments, the number of prediction steps can be denoted as L, and the number of base models as M. For example, at a 10-minute time resolution, a 4-hour prediction corresponds to 24 prediction steps, a 12-hour prediction corresponds to 72 prediction steps, and a 24-hour prediction corresponds to 144 prediction steps; in this embodiment, the number of base models M is 3.

[0056] Weight logarithm matrix: Define a trainable parameter matrix W of shape L×M, where the element W[t,m] represents the unnormalized logit of the weights of the m-th base model at the t-th prediction step. Initialize all elements to zero, which is equivalent to initial equal weights.

[0057] This design of the weight parameter matrix has a significant advantage: the number of parameters is extremely small, only L×M. Taking 4-hour prediction as an example, only 24×3=72 trainable parameters are needed; 12-hour prediction requires 216; and 24-hour prediction requires only 432. Such a small number of parameters brings two direct technical benefits: first, the fusion layer is almost impossible to overfit, and can be trained stably even with limited validation data; second, the training speed is extremely fast, and the training of the fusion layer can usually be completed in just a few minutes to a dozen minutes, with very low computational resource overhead.

[0058] Step S202: The weight parameters corresponding to each prediction step are gradually normalized using a normalized exponential function.

[0059] Specifically, the normalization exponential function generally refers to the softmax function, which is a function that transforms any real vector into a non-negative probability distribution that sums to 1. In this step, the normalization exponential function operation is performed once for each prediction step to obtain the combination coefficients of each base model in that prediction step. The specific calculation method is shown in equation (1): Equation (1) in, This represents the combination coefficient of the m-th base model at the t-th prediction step, which is also the contribution ratio of each base model at the t-th step. t indicates the nth time step in the prediction sequence, and m is the index of the base model. This represents the unnormalized logit of the weights of the m-th base model at the t-th prediction step. It is an exponential function (with the natural constant e as the base). For any prediction step t, all its corresponding combination coefficients A[t,:] are non-negative and their sum is 1.

[0060] It's important to clarify that the progressive normalization in this method means that each prediction step performs a normalization operation independently, without affecting each other. In other words, the weight normalization result in step 1 will not affect the weights in step 2; each step has its own independent set of weights. It's not about assigning a uniform set of weights to the entire prediction sequence, but rather assigning a separate set of weights to each prediction step.

[0061] Step S203: According to the combination coefficients corresponding to each prediction step, the predicted values ​​of each deep learning base model at the corresponding prediction step are weighted and summed step by step to obtain the fused prediction sequence.

[0062] Specifically, let p be the predicted sequence output by the m-th base model for a certain sample. m (Length L), then the t-th step of the fused prediction sequence ŷ is as shown in equation (2): Equation (2) in, This represents the fused prediction value at the t-th prediction step. This represents the combination coefficients of the m-th base model at the t-th prediction step. This represents the predicted value of the m-th base model at the t-th prediction step. It's important to note that each prediction step uses only the combined coefficients of its own row for weighting calculations; the prediction steps do not affect each other. In other words, each prediction step uses only the weights of its own row, and they do not influence each other.

[0063] The technical principle behind this time-step weighted summation design is that the optimal model combination differs at different prediction distances in multi-step prediction tasks. In near-term prediction steps, models adept at modeling local details typically perform better and should receive higher weights. In far-term prediction steps, the errors of each model increase, and no single model can consistently dominate. Therefore, a relatively balanced combination of multiple models is needed to improve overall accuracy through error cancellation. Traditional fusion methods that use a single set of weights across the entire sequence cannot capture this difference. The time-step fusion mechanism of this invention adaptively learns the optimal combination weights for each prediction step, thereby achieving better results across the entire prediction sequence.

[0064] Furthermore, since the combination coefficients can be directly read, the fusion mechanism of this invention inherently possesses good interpretability. After training, simply printing out the combination coefficient matrix A provides a clear view of "which step mainly relies on which model, and whether the long-term step is a multi-model equilibrium," without the need for additional attribution analysis methods. This interpretability is highly valuable for practical engineering applications: on the one hand, it facilitates understanding the model's decision-making logic, enhancing user trust in the model; on the other hand, when abnormal predictions occur, it allows for quick identification of which base model is malfunctioning, facilitating debugging and troubleshooting.

[0065] In addition to the aforementioned forward fusion process, the training input for the attention-stacking fusion layer of this invention is the predicted sequences output by each base model on the validation set, and the supervision signal is the real observation label. This differs from traditional stacked generalization methods, which typically use the predictions of the base models on the training set to train the meta-learner. This invention uses validation set predictions as training input because the predictions of the base models on the training set often exhibit overfitting, especially for long short-term memory models with a large number of parameters, where the error on the training set is often much smaller than its true generalization error. Using validation set predictions to train the fusion layer, since the validation set data is not involved in the training of the base models, better reflects the true generalization ability of the base models, thus making the weights learned by the fusion layer more reliable and its performance on real data more stable. This training strategy effectively avoids the contamination of the fusion weights by overfitting of the base model training set.

[0066] In terms of training details, the fusion layer can use the Huber loss function to calculate the prediction error and employ the Adam adaptive moment estimator for gradient descent iterations, with the learning rate typically set to 0.01. The Huber loss function combines the advantages of mean squared error and mean absolute error, is insensitive to outliers, and is more robust than the simple mean squared error loss function; the Adam adaptive moment estimator can adaptively adjust the learning rate of each parameter, has a fast convergence speed, and good training stability.

[0067] Furthermore, to prevent the fusion layer from overfitting on the validation set, a small subset of the validation set predictions can be further divided into an internal validation set, which can be used in conjunction with an early stopping strategy. Training is stopped early when the loss on the internal validation set stops decreasing for several consecutive rounds, thus preventing the fusion layer from overfitting to the validation set data.

[0068] The entire solution exposes only two time-domain related parameters: the input window length and the prediction window length. These two parameters are passed down the process: affecting the sliding window design, the output dimension of the base model's prediction head, and the number of rows in the fusion layer's weight matrix. Everything else (feature fields, network backbone, training process) remains consistent. Therefore, changing the prediction duration only requires changing these two numbers, without needing to redesign the model.

[0069] In summary, the attention stacking fusion layer of the present invention achieves accurate adaptive fusion of different prediction steps in multi-temporal prediction of total solar radiation through the above-mentioned weight parameter matrix construction, normalization of the normalized exponential function, weighted summation step by step, and training strategy based on the validation set prediction, which significantly improves the accuracy and reliability of fusion prediction.

[0070] Figure 3 This is a flowchart illustrating a multi-temporal-domain prediction method for time-step total solar radiation according to another embodiment of the present invention, as shown below. Figure 3As shown, the method of this invention is divided into multiple layers from top to bottom according to the data flow direction, namely: multi-source data input, data preprocessing, sliding window sequence construction, site-based spatial partitioning, model training, comprehensive evaluation, and final output. These layers are sequentially connected to form a complete prediction pipeline. The functions, principles, and technical effects of each layer are described in detail below with reference to the accompanying drawings.

[0071] Step 1: Obtain the data source.

[0072] To comprehensively capture all meteorological factors affecting total solar radiation, this invention integrates multi-source heterogeneous data, specifically including three major data sources: (1) Satellite remote sensing data: including 6 channels of shortwave albedo and 10 channels of infrared brightness temperature. Among them, shortwave albedo is used to characterize cloud top reflection characteristics and aerosol optical thickness, while infrared brightness temperature reflects cloud top height and atmospheric water vapor content. The combination of the two can provide information on cloud state and atmospheric vertical structure over a wide range.

[0073] (2) Ground observation station data: This includes seven ground meteorological elements from multiple observation stations (e.g., 91 meteorological bureau stations): air temperature, wind direction, wind speed, relative humidity, air pressure, surface temperature, and visibility, as well as historical total solar irradiance (GHI, Global Horizontal Irradiance, i.e., the total solar radiation power received per unit horizontal surface). Historical GHI, as an autoregressive input, can provide the model with time-series dependence information on the radiation sequence itself, playing a crucial role in capturing recent radiation change trends.

[0074] (3) Observational geometric angle data: including satellite zenith angle (SAZ), satellite azimuth angle (SAA), solar zenith angle (SOZ), and solar azimuth angle (SOA). The solar zenith angle (SOZ) and solar azimuth angle (SOA) are calculated from the station's latitude and longitude and the observation time; the satellite zenith angle (SAZ) and satellite azimuth angle (SAA) depend on the satellite's orbital position and the sensor's scanning geometry. These parameters determine the length of the radiation transmission path and are key parameters affecting atmospheric radiation attenuation and satellite observation inversion.

[0075] The three types of data mentioned above come from different sources, but they are all aligned to the same time resolution (e.g., 10 minutes) and used as input fields for the model feature table.

[0076] Step 2, data preprocessing.

[0077] Because different data sources have different collection methods, units, and formats, directly inputting them into the model can affect the model training effect. Therefore, this embodiment preprocesses the data before feature input, mainly including: (1) Quality control: Remove abnormal data whose observed values ​​exceed the reasonable physical range, such as observed values ​​that exceed the reasonable physical range, abnormal values ​​caused by instrument failure, and jump values ​​that obviously do not conform to the time series pattern.

[0078] (2) Missing value imputation: Meteorological observation data often contains missing values ​​due to instrument maintenance, communication failures, severe weather, etc., and imputation is necessary to ensure the continuity of the data sequence. The specific strategy is as follows: fields with low missing values ​​are imputed forward or backward, that is, they are filled using observations from adjacent time points; fields that still have missing values ​​are filled using the column mean, that is, the overall average value of the field is used. Reasonable missing value imputation can ensure the continuity of the data sequence and provide a complete data foundation for subsequent sliding window construction.

[0079] (3) Feature merging: Satellite data, ground data and geometric angle data are aligned and spliced ​​according to timestamp to form a unified feature table, ensuring the spatiotemporal consistency of features in each dimension at the same time step.

[0080] (4) Standardization: A standardization method (e.g., StandardScaler) is used to fit and transform the dynamic features, static features, and predicted labels in the feature table respectively, so that all features are within a similar scale range. This operation can effectively improve the convergence speed and stability of subsequent base model training.

[0081] After preprocessing, each station corresponds to 33-dimensional dynamic features and 3-dimensional static features. The dynamic features include satellite albedo, infrared brightness temperature, geometric angle, ground meteorological elements, time coding, and historical target values; the static features include latitude, longitude, and altitude, reflecting the spatial attributes of the station.

[0082] Step 3: Construct the sliding window sequence.

[0083] Specifically, the input window length and prediction window length are first determined based on the target prediction time domain. Then, the window slides across the feature sequence according to a set step size, extracting a segment of the input sequence and its corresponding output sequence each time to form a training sample. This invention supports multiple prediction time domains, such as 4 hours, 12 hours, and 24 hours, with different window configurations corresponding to different time domains. Taking a 10-minute resolution as an example: 4-hour prediction uses 144 steps (24 hours) of input corresponding to 24 steps of output, with a window step size of 6; 12-hour prediction uses 288 steps (48 hours) of input corresponding to 72 steps of output, with a window step size of 12; 24-hour prediction also uses 288 steps of input corresponding to 144 steps of output, with a window step size of 18.

[0084] Specifically, this embodiment sets the input window length to cover at least two complete diurnal cycles. This is because total solar radiation exhibits a significant diurnal cyclical variation, with features such as sunrise, sunset, and noon peak showing clear periodicity. The model needs to observe at least two complete day-night cycles to fully learn this cyclical pattern and thus more accurately predict future radiation changes.

[0085] By using a sliding window sequence, continuous feature sequences can be transformed into input-output sample pairs that the model can directly process, providing a standardized data foundation for subsequent model training; at the same time, by flexibly configuring the window length, it can support the needs of different prediction time domains and has good versatility.

[0086] Step 4: Spatial partitioning based on the site.

[0087] To accurately evaluate the model's spatial generalization ability under different climate zones and altitudes, this embodiment of the invention does not employ random or chronological data partitioning, but instead uses a site-based spatial partitioning strategy. Specifically, a total of 91 observation sites are stratified and sampled according to multiple climate zones and altitude gradients, dividing them into a training set (e.g., 73 sites, approximately 80%) and a validation set (e.g., 18 sites, approximately 20%). The sites in the training and validation sets do not overlap. The validation sites cover various climate types, including temperate monsoon, temperate continental, subtropical monsoon, tropical monsoon, plateau mountain, and subtropical alpine, with altitudes ranging from 2 meters to 3649 meters. This partitioning method avoids the problem of inflated validation accuracy caused by data from the same site appearing in both the training and validation sets, accurately reflecting the model's predictive performance at unfamiliar sites, and ensuring the objectivity and reliability of the model performance evaluation.

[0088] Step 5, Model Training (including base model training and stacked fusion layer training).

[0089] This step is specifically divided into two levels: base model training and fusion layer training.

[0090] First, three structurally distinct deep learning base models are trained in parallel. This structural dissimilarity ensures that the different models have different inductive biases, thus providing complementary information during subsequent fusion. The three base models specifically include: (1) CNN-BiLSTM-Attention model: It consists of a dilated convolutional neural network (CNN), a bidirectional long short-term memory network (BiLSTM), and a multi-head self-attention cascade. The cascade of the three can take into account local feature extraction, long-range dependency modeling and adaptive information aggregation, and usually performs well in short-to-medium-term prediction.

[0091] (2) Transformer model: It adopts a 4-layer 8-head self-attention encoder structure. Its self-attention mechanism allows any two time steps in the sequence to interact directly. It has a global vision and is good at capturing long-distance dependencies.

[0092] (3) BiLSTM-Attention model: It consists of bidirectional LSTM and Bahdanau additive attention. The number of parameters is relatively small, and the error decay is relatively stable in the long time domain. It serves as an ablation baseline and control.

[0093] The three base models do not exchange information and are trained independently. After each base model is trained, its prediction sequence for the validation site set is input into the Stacking-Attention fusion layer (i.e., the attention stacking fusion layer).

[0094] The attention-stacking fusion layer no longer processes the original meteorological features during training. Instead, it uses only the prediction sequences of each base model for the validation site set as input and the actual total solar radiation values ​​of the validation site set as supervision signals for training. Specifically, it constructs an L x M column trainable weight parameter matrix, independently maintaining a set of weights for each base model for each prediction step in the output sequence. The weights for each prediction step are normalized using a normalized exponential function (softmax), and finally, the predicted values ​​of each base model are weighted and summed for each prediction step. This fusion layer uses validation set predictions instead of training set predictions as training input to effectively avoid contamination caused by overfitting of the base model training set. If training set predictions are used, the fusion weights will be misled by the severely overfitted base models, leading to degradation in real-world applications. Furthermore, this fusion layer has a very small number of parameters (e.g., only 72 parameters for a 4-hour prediction), making overfitting virtually impossible under limited validation data. After training, it can directly output the weight distribution for each prediction step, naturally possessing strong interpretability.

[0095] Step Six: Comprehensive Evaluation.

[0096] To comprehensively and objectively evaluate model performance, this embodiment of the invention simultaneously calculates three types of metrics on both the fusion prediction and the predictions of each base model: (1) Global indicators: including root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²). 2 The relative root mean square error (rRMSE) is used to reflect the overall error level of the entire predicted sequence.

[0097] (2) Daytime indicators: These include the root mean square error (RMSE_day), the mean absolute error (MAE_day), and the normalized mean absolute percentage error (nMAPE_day). Since total solar radiation approaches zero at night, the daytime indicators are lowered by the nighttime samples; only daytime indicators with a label mean greater than 5 W / m² are retained. 2 The sample size can more accurately reflect the model's true predictive ability during the effective daytime power generation period.

[0098] (3) Sub-indicators: Statistics are compiled for 18 verification sites, four seasons (spring, summer, autumn and winter) and each prediction step to determine the best or worst performance of the positioning model under what climatic conditions, in what season and at what prediction distance.

[0099] Step 7: Output the optimal result.

[0100] Following the above process, the final output is a fused prediction sequence generated by the Stacking-Attention fusion layer. In actual testing, this fusion model achieved the best performance metrics compared to any single base model in the three prediction time domains of 4 hours, 12 hours, and 24 hours. Furthermore, as the prediction time domain lengthens, the improvement of the fusion model compared to the best single model is even greater, indicating that the advantages of multi-model complementarity are particularly significant in long-term predictions. This verifies the effectiveness and reliability of the method of this invention in improving the accuracy of multi-step total solar radiation prediction.

[0101] The effects achieved by the above steps mainly include: (1) In multi-step solar total radiation prediction, each prediction step adaptively obtains a set of optimal combined weights for each base model, instead of using a single weight shared by the entire sequence. (2) Make the fusion mechanism itself have enough trainable parameters and strong enough constraints so that it will not overfit when the validation data is limited, and can directly read the contribution ratio of each base model in each prediction step, which is convenient for review and analysis. (3) Avoid the pollution of fusion weights by overfitting of the base model training set, so that the fusion remains effective on real unseen data; (4) Use the same methodology to cover multiple prediction time domains such as 4 hours, 12 hours, and 24 hours, and switch only by configuring parameters without redoing the model structure; (5) Ensure that the method has the ability to be truly evaluated in terms of its spatial generalization ability in different climate zones, different altitudes, and different observation conditions.

[0102] Ultimately, the embodiments of the present invention successfully achieved high-precision adaptive prediction of total solar radiation in multiple time domains by ensuring the quality of input data, simulating different time-domain periodicity, realistically evaluating spatial generalization ability, utilizing the complementary advantages of base models, and adaptive fusion step by step.

[0103] To further illustrate the method of the present invention, a specific embodiment is provided for description: First, the key technical parameters of this embodiment are shown in Table 1: Table 1 Key Technical Parameters Following the key technical parameters shown in Table 1, the method of this invention was executed, and the final results are shown in Table 2. The results indicate that the Stacking-Attention model achieved optimal performance in all cases. The R² for prediction after 4 hours reached 0.8793, and the RMSE was 93.98 W / m². 2 As the prediction time domain was extended from 4 hours to 24 hours, R² decreased from 0.8793 to 0.8187, while RMSE increased from 93.98 to 115.29 W / m². 2 This reflects the general rule that the longer the prediction time domain, the greater the uncertainty in total solar radiation prediction.

[0104] Table 2 Model Performance Results Because the total solar radiation data includes a large number of zero-value samples at night (accounting for approximately 41% of the total sample), the all-time indicators (RMSE, R...) 2 The predictability of daytime data is negatively impacted by the "predictability" factor at night, failing to accurately reflect the model's actual predictive ability. Therefore, a daytime metric (GHI > 5 W / m² threshold for filtering nighttime samples) is introduced to more accurately evaluate daytime predictive performance. An example of an optional daytime metric is shown in Table 3. Table 3 Model Daytime Indicators As shown above, the RMSE during the day is generally higher than the RMSE at all times (e.g., in the 4-hour fusion model: 122.61 W / m during the day). 2 vs 93.98W / m 2 This is because daytime radiation values ​​are higher and more variable. The normalized mean absolute percentage error (nMAPE_day) increases with the forecast time domain: 23.61% for 4 hours, 29.85% for 12 hours, and 33.63% for 24 hours, indicating that the cumulative effect of uncertainty in long-term forecasts is more significant during the high-radiation period of daytime.

[0105] Subsequently, we conducted a site-by-site improvement analysis on the 18 verification sites and selected 6 typical sites for display. The results are shown in Table 4.

[0106] Table 4. Station-by-station improvement analysis of the fusion model As shown in Table 4, the fusion model achieved R in 5 out of the 6 sites. 2 The improvement was only slightly diminished at Mount Emei (a decrease of 0.0065). The fusion gain was most significant at sites of moderate difficulty (e.g., Lhasa +0.0074, Ganzhou +0.0050), while its effect was limited at the best-performing (Tianjin +0.0009) and worst-performing (Mount Emei -0.0065) sites. This phenomenon can be explained as follows: for easily predictable sites, the single model is already close to its prediction limit, and the improvement potential of fusion is limited; for extremely difficult sites (e.g., Mount Emei), the prediction errors of the three base models are all large and may be in the same direction, making effective error elimination impossible through fusion.

[0107] In practical use, the improvements brought about by the method of the present invention can be seen from the following aspects through comparison: 1. Progressive fine-grained fusion outperforms the best single model across all time domains, with greater gains over longer time domains. On 18 validation sites covering nine geographical regions (Northeast, North China, East China, Central China, South China, Southwest, Northwest, Inner Mongolia, and the Qinghai-Tibet Plateau), spanning latitudes from 18°N to 47°N and altitudes from 2.5m to 3648.7m, the fusion model's 4-hour predicted R... 2 Approximately 0.879, RMSE approximately 93.98 W / m 2 ; 12 hours R 2 Approximately 0.840; 24-hour R 2 Approximately 0.819. R0 for the relative equal-weighted average. 2 The gains are 0.53% / 1.72% / 2.61% – the gains increase as the time horizon advances. This aligns with the design expectation: in the long term, no single model will dominate, and the potential for multi-model complementarity is maximized.

[0108] 2. The fusion mechanism itself has very few parameters. The entire layer has only L×M parameters. A 4-hour prediction yields only 72 parameters, an order of magnitude smaller than the prediction head of any base model. This order of magnitude is crucial when validation data is limited—it prevents the fusion layer from overfitting the validation set.

[0109] 3. The weights can be read directly. After training, printing matrix A immediately reveals that the CBA weights are largest in the near prediction steps (such as t+1) (typically around 0.77), and as t progresses, the weights of the three base models gradually level off. This level of interpretability is directly useful for attributing anomaly predictions and for reviewing errors in the runtime environment.

[0110] 4. The effectiveness of meta-training on the validation set can be verified. In control experiments, if the meta-learner is trained on the training set predictions, the weights will clearly favor the BiLSTM-type model with the most severe overfitting, and will degenerate on the validation set; only by using validation set predictions for training can the fusion model stably outperform the single model.

[0111] 5. The engineering significance of a unified multi-time-domain framework is direct. The 4-hour, 12-hour, and 24-hour time domains share the same training and evaluation scripts at the code level, distinguished only by the configuration. Maintaining a single model can simultaneously support intraday scheduling, daily planning, and next-day forecasting, eliminating the repetitive work of maintaining separate models for each time domain.

[0112] 6. Spatial generalization ability can be realistically evaluated. Six validation sites were not included in the training. The fusion model achieved R0 relative to the best single model on five of these sites. 2 The improvement was most significant for sites with moderate prediction difficulty (such as Lhasa +0.0074, Ganzhou +0.0050); however, for extremely difficult sites (such as Mount Emei, with more than 300 foggy days per year), even with large errors in individual models, fusion could not improve the results. This is where the -R value in this solution is not significantly improved. 2 The return to Emei Mountain was slightly lower by about 0.0065, indicating that the fusion of the benefit-dependent base model has reached a certain basic accuracy.

[0113] 7. Low resource overhead. The fusion layer has no convolutions or loops. After the base model is trained, the fusion layer can usually be trained in a few minutes to a dozen minutes on a normal workstation; the inference stage only involves matrix multiplication, which is negligible relative to the base model itself.

[0114] The flowchart provided in this embodiment is not intended to indicate that the operations of the method will be performed in any particular order, or that all operations of the method are included in every case. Furthermore, the method may include additional operations. Within the scope of the technical concept provided by the method in this embodiment, additional variations can be made to the above method.

[0115] It should be understood that in some embodiments, the components may be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods may be implemented using software or firmware stored in memory and executed by a suitable instruction execution system.

[0116] This embodiment also provides a computer program product 10, a computer-readable storage medium 20, and a computer device 30. Figure 4 This is a schematic diagram of a computer program product 10 according to an embodiment of the present invention. Figure 5 This is a schematic diagram of a computer-readable storage medium 20 according to an embodiment of the present invention. Figure 6This is a schematic diagram of a computer device 30 according to an embodiment of the present invention. The computer program product 10 includes a computer program 11, which, when executed by the processor 32, implements the steps of any of the above-described time-step multi-temporal prediction methods for total solar radiation. A computer-readable storage medium 20 stores the computer program 11 thereon, which, when executed by the processor 32, implements the steps of any of the above-described time-step multi-temporal prediction methods for total solar radiation. The computer device 30 may include a memory 31, a processor 32, and the computer program 11 stored in the memory 31 and running on the processor 32.

[0117] The computer program 11 used to perform the operations of this invention may be assembly instructions, Instruction Set Architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, state setting data, integrated circuit configuration data, or source code or object code written in any combination of one or more programming languages ​​and procedural programming languages. The computer program 11 may execute 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 the latter case, the remote computer may be connected to the user's computer via any type of network, including a Local Area Network (LAN) or Wide Area Network (WAN), or may be connected to an external computer (e.g., via the Internet using an Internet service provider). In some embodiments, to perform aspects of this invention, electronic circuits, including, for example, programmable logic circuits, Field-Programmable Gate Arrays (FPGAs), or Programmable Logic Arrays (PLAs), may execute computer-readable program instructions to personalize the electronic circuits by utilizing state information from computer-readable program instructions.

[0118] For the purposes of this embodiment, computer program product 10 is a related product containing computer program 11. For the purposes of this embodiment, computer-readable storage medium 20 is a tangible device capable of holding and storing computer program 11, and can be any device capable of containing, storing, communicating, propagating, or transmitting program 11 for use by or in conjunction with an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable storage medium 20 include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable optical disc read-only memory (CD-ROM), digital versatile disc (DVD), memory stick, floppy disk, mechanical encoding device, and any suitable combination thereof.

[0119] Computer device 30 can be, for example, a server, desktop computer, laptop computer, tablet computer, or smartphone. In some examples, computer device 30 can be a cloud computing node. Computer device 30 can be described in the general context of computer system executable instructions (such as program modules) executed by a computer system. Typically, program modules can include routines, programs, object programs, components, logic, data structures, etc., that perform specific tasks or implement specific abstract data types. Computer device 30 can be implemented in a distributed cloud computing environment where tasks are performed by remote processing devices linked through a communication network. In a distributed cloud computing environment, program modules can reside on local or remote computing system storage media, including storage devices.

[0120] Computer device 30 may include a processor 32 adapted to execute stored instructions and a memory 31 that provides temporary storage space for the operation of said instructions during operation. The processor 32 may be a single-core processor, a multi-core processor, a computing cluster, or any other configuration. The memory 31 may include random access memory (RAM), read-only memory, flash memory, or any other suitable storage system.

[0121] Computer device 30 may also include a network adapter / interface and an input / output (I / O) interface. The I / O interface allows external devices that can be connected to the computer device to input and output data. The network adapter / interface provides communication between the computer device and a network, typically represented as a communication network.

[0122] Therefore, those skilled in the art should recognize that although numerous exemplary embodiments of the present invention have been shown and described in detail herein, many other variations or modifications conforming to the principles of the present invention can be directly determined or derived from the disclosure of the present invention without departing from the spirit and scope of the invention. Thus, the scope of the present invention should be understood and construed as covering all such other variations or modifications.

Claims

1. A multi-temporal domain prediction method for total solar radiation step-by-step, comprising: Meteorological observation data from multiple data sources are acquired and preprocessed to obtain model feature data; Multiple sliding window sequences are constructed based on the model feature data and multiple preset target prediction time domains; Multiple deep learning base models with distinct structures are trained using the sliding window sequence. These deep learning base models are used to independently output a base prediction sequence containing multiple prediction steps. The basic prediction sequences output by multiple deep learning base models are fed into an attention stacking fusion layer for fusion processing to obtain a fused prediction sequence. The attention stacking fusion layer maintains an independent set of weights for each prediction step in the basic prediction sequence for each deep learning base model, and is used to receive the basic prediction sequences output by each deep learning base model. After normalizing the weights of each prediction step, the predicted values ​​of each deep learning base model are weighted and summed according to the prediction step. The multi-step total solar radiation prediction results in multiple time domains are output based on the fused prediction sequence.

2. The method for multi-temporal-domain prediction of total solar radiation step-by-step according to claim 1, wherein, The step of feeding the basic prediction sequences output by multiple deep learning base models into an attention stacking fusion layer for fusion processing to obtain a fused prediction sequence includes: Construct a weight parameter matrix, wherein the number of rows in the weight parameter matrix is ​​equal to the number of prediction steps and the number of columns is equal to the number of deep learning base models. The elements in the weight parameter matrix represent the unnormalized logarithm of the weights of the deep learning base models at the corresponding prediction steps. The weight parameter matrix is ​​initialized to zero. The weight parameters corresponding to each prediction step are gradually normalized using a normalized exponential function to obtain the combination coefficients corresponding to the prediction step. The combination coefficients are non-negative and sum to 1. The combination coefficients represent the contribution ratio of each deep learning base model in the prediction step. According to the combination coefficients corresponding to each prediction step, the predicted values ​​of each deep learning base model at the corresponding prediction step are weighted and summed step by step to obtain the fusion prediction sequence.

3. The multi-temporal domain prediction method for total solar radiation step-by-step according to claim 1, wherein, The data sources include: satellite remote sensing data, ground observation station data, and observation geometric angle data; The satellite remote sensing data includes: multi-channel shortwave albedo and multi-channel infrared brightness temperature; The ground observation station data includes: air temperature, wind direction, wind speed, relative humidity, air pressure, surface temperature, visibility, and historical total solar radiation from multiple observation stations; The observed geometric angle data includes: solar zenith angle, solar azimuth angle, satellite zenith angle, and satellite azimuth angle; The sliding window sequence includes an input window sequence, a prediction window sequence, and a sliding window step size; The deep learning base models include a convolutional neural network-bidirectional long short-term memory network-attention cascade model, a self-attention encoder model, and a bidirectional long short-term memory network-additive attention model.

4. The multi-temporal domain prediction method for total solar radiation step-by-step according to claim 3, wherein, The step of constructing multiple sets of sliding window sequences based on the model feature data and multiple preset target prediction time domains includes: Based on the preset multiple target prediction time domains, determine the input window length, prediction window length, and sliding window step size corresponding to each target prediction time domain; Based on the input window length, the prediction window length, and the sliding window step size, the input window sequence and the corresponding prediction window sequence are extracted from the model feature data, and the input window length is set to cover at least two complete daily cycles.

5. The method for multi-temporal-domain prediction of total solar radiation step-by-step according to claim 1, wherein, Before the step of feeding the basic prediction sequences output by the multiple deep learning base models into the attention stacking fusion layer for fusion processing, the method further includes: The meteorological observation data is divided into a training set and a validation set, and the stations in the training set and the validation set are different from each other. The training set is used to train the deep learning base model, and the validation set is used to evaluate the generalization ability of the model.

6. The method for multi-temporal-domain prediction of total solar radiation step-by-step according to claim 5, wherein, The training steps for the attention stacking fusion layer include: The predicted sequences output by each of the deep learning base models on the validation set are obtained as training inputs, and the real observations are used as supervision signals. Determine the loss function and optimizer to be used for training; Based on the training input and the supervision signal, the weight parameters of the attention stack fusion layer are updated; An internal validation set is partitioned from the validation set predictions and an early stopping strategy is employed to prevent overfitting of the attention stacked fusion layer.

7. The method for multi-temporal-domain prediction of total solar radiation step-by-step according to claim 1, wherein, After the step of feeding the basic prediction sequences output by the multiple deep learning base models into the attention stacking fusion layer for fusion processing, the method further includes: The prediction performance indices of the fused prediction sequence and each basic prediction sequence are calculated. The prediction performance indices include global indices, daytime indices, and sub-indices. The global indices are used to reflect the overall error of the entire prediction sequence. The daytime indices are used to remove the influence of zero-value samples at night and reflect the daytime prediction performance. The sub-indices are statistically analyzed according to the verification station, season, and prediction step, and are used to locate the prediction performance under different climatic conditions and different prediction distances.

8. The method for multi-temporal-domain prediction of total solar radiation step-by-step according to claim 1, wherein, The preprocessing steps include: The meteorological observation data is subjected to quality control to remove abnormal data; Missing values ​​were filled into the meteorological observation data. The meteorological observation data are merged according to timestamps to form a unified feature table; The dynamic features, static features, and predicted labels in the feature table are standardized to obtain the model feature data.

9. A computer program product comprising a computer program that, when executed by a processor, implements the steps of the time-step multi-temporal-domain prediction method for total solar radiation as described in any one of claims 1 to 8.

10. A computer device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor, when executing the computer program, implements the steps of the multi-temporal-domain prediction method for time-step total solar radiation according to any one of claims 1 to 8.

Images