Solar radiation space-time prediction method and system based on physical information constraint and neural network

Abstract

The invention discloses a solar radiation space-time prediction method and system based on physical information constraint and a neural network, and the method comprises the steps: collecting multi-dimensional time sequence meteorological data, extracting high-dimensional time sequence dynamic features, converting geographic space data into a fuzzy set, and carrying out the defuzzification of the fuzzy set through an inference rule, thereby obtaining geographic space features; and a gating mechanism is adopted to realize deep fusion of the space-time features to generate high-dimensional space-time fusion features. In a model training stage, an energy conservation equation is introduced into an optimization process, a physical residual error is constructed by calculating a time derivative and a space derivative of a predicted value, a physical constraint total loss function is formed in combination with a data loss item, and model parameters are updated by using a gradient descent method. According to the method, the accuracy and reliability of a prediction result are remarkably improved while the calculation efficiency is ensured, and the method is particularly suitable for solar radiation prediction under complex meteorological conditions; according to the method, abnormal prediction caused by data noise can be effectively corrected, and a solution with physical rationality and data adaptability is provided for the fields of solar resource evaluation, photovoltaic power generation power prediction and the like.

Description

Technical Field

[0001] This invention relates to the field of solar radiation prediction technology, specifically to a method and system for spatiotemporal prediction of solar radiation based on physical information constraints and neural networks. Background Technology

[0002] Solar radiation is the source of the vast majority of energy on Earth's surface, and its spatiotemporal distribution directly affects photovoltaic power generation efficiency, agricultural production activities, climate modeling, and the operation and scheduling of renewable energy systems. Therefore, accurate spatiotemporal prediction of solar radiation is not only a crucial foundation for promoting the efficient development and utilization of solar energy resources, but also a key link in ensuring the stable operation of the power system, optimizing the energy structure, and supporting meteorological and environmental science research. Currently, research in this field mainly follows two technical paths: one is traditional statistical and machine learning methods, and the other is time series modeling methods represented by deep learning.

[0003] Traditional statistical models, such as the Autoregressive Moving Average (ARMA) model and the Exponential Smoothing (ES) method, have been widely used in solar radiation prediction tasks due to their mature theory, computational simplicity, and ease of implementation. These methods are typically based on the stationarity and linear correlation of time series data, modeling the statistical characteristics of historical observation data to extrapolate future radiation trends. However, the formation and variation of solar radiation are constrained by the Earth-Sun motion, geographical factors (such as latitude, longitude, and altitude), and complex meteorological conditions (such as cloud cover, aerosols, humidity, and temperature). These factors exhibit highly nonlinear and coupled relationships, resulting in strong fluctuations, intermittency, and nonstationarity in solar radiation. Traditional statistical models fall short in handling complex nonlinear characteristics and medium- to long-term prediction tasks, especially in variable climate or extreme weather scenarios, where prediction accuracy significantly decreases, making it difficult to meet the needs of refined applications.

[0004] With the development of artificial intelligence technology, especially deep learning, recurrent neural networks (RNNs) and their improved structures—such as long short-term memory networks (LSTM) and gated recurrent units (GRUs)—have shown significant advantages in time series forecasting. Although deep learning-based methods for solar radiation forecasting have made considerable progress in recent years, they still face the following key challenges and technical bottlenecks in practical deployment and large-scale promotion, which urgently need further resolution: First, insufficient representation of spatial heterogeneity is a major weakness of current models. Most deep learning models focus on single-point time series modeling, treating each observation station as an independent entity and ignoring the continuity and inherent correlation of solar radiation across geographic space. This approach limits the model's ability to generalize across regions or the entire country, especially in unobserved or marginal areas where prediction performance declines significantly. Second, a lack of physical consistency is a common problem for purely data-driven models. These models rely on the statistical distribution reflected in the training data, lacking the guidance and constraints of physical knowledge. In areas with sparse data, extreme weather events, or new scenarios, they may output predictions that violate physical laws, such as abnormally high nighttime predictions or misinterpreting cloudy days as strong radiation, thus affecting the accuracy of photovoltaic power plant scheduling strategies and energy balance assessments. Furthermore, strong data dependence and high computational costs are also significant issues. Deep neural network models require large amounts of high-quality historical data for training, while the regional imbalances and inconsistent formats of solar radiation data increase the difficulty of data acquisition and processing. Meanwhile, to enhance modeling capabilities, model structures are becoming increasingly complex, leading to a surge in the number of parameters, high computational resource consumption, long training times, and high deployment costs, limiting their widespread application in small and medium-sized enterprises or edge devices. Finally, the challenges of multi-source information fusion should not be underestimated. The formation and evolution mechanisms of solar radiation are complex and diverse, involving multiple factors such as astronomical periodicity, geographic spatial distribution, instantaneous meteorological changes, and solar atmospheric activity. Therefore, single-type information is often insufficient to support high-precision predictions; it is essential to effectively integrate time-series information, spatial information, physical knowledge, and other auxiliary data. However, existing models are still rudimentary in multi-source heterogeneous data fusion, with overly simplistic fusion mechanisms that may lead to information redundancy, mutual interference, or weakened weights of key factors, thereby affecting overall model performance.

[0005] Despite significant progress made by deep learning in solar radiation prediction, existing technologies still face key challenges such as insufficient spatial modeling capabilities, lack of physical consistency, strong data dependence, and low efficiency in fusing multi-source information. These issues limit the accuracy, generalization ability, and practical application value of prediction models, and urgently require solutions through innovative methods and technologies to drive the development of solar radiation prediction technology towards higher accuracy, stronger generalization ability, and better physical consistency. Summary of the Invention

[0006] To address the problems existing in the prior art, this invention provides a spatiotemporal prediction method and system for solar radiation based on physical information constraints and neural networks. This overcomes the shortcomings of existing solar radiation prediction methods in terms of spatial feature representation, physical consistency assurance, multi-source information fusion, and prediction accuracy and generalization ability, thereby achieving accurate spatiotemporal prediction of solar radiation across the entire country.

[0007] This invention is achieved through the following technical solution: In a first aspect, this application provides a method for spatiotemporal prediction of solar radiation based on physical information constraints and neural networks, comprising the following steps: Step 1: Obtain the high-dimensional time-series dynamic characteristics of multi-dimensional time-series meteorological data; Step 2: Collect geospatial data and convert it into a fuzzy set. Use inference rules to perform fuzzy inference on each geographic attribute in the fuzzy set, and defuzzify the inference results to obtain geospatial features. Step 3: Use a gating mechanism to fuse high-dimensional temporal dynamic features and geospatial features to obtain high-dimensional spatiotemporal fusion features; Step 4: Train the data-driven model. The training method is as follows: The high-dimensional spatiotemporal fusion features are input into the data-driven model to obtain the current radiation prediction value. The physical relationship between meteorological data and the current radiation prediction value is determined according to the energy conservation equation. The physical residual is determined according to the radiation prediction value and the physical relationship. The physical residual is combined with the data loss term to obtain the physical constraint total loss function. The parameters of the data-driven model are updated using the physical constraint total loss function and iteratively trained until the physical constraint total loss function is minimized, resulting in a trained data-driven model. Solar radiation values ​​are then predicted based on the trained data-driven model.

[0008] Preferably, step 1 involves using a multi-dimensional time-series dynamic extraction module to collect multi-dimensional time-series meteorological data and obtain high-dimensional time-series dynamic features of the multi-source time-series meteorological data, including: Multidimensional time-series meteorological data are collected and combined with the sliding window segmentation method to divide the multidimensional time-series meteorological data into meteorological data sequence segments. Local time-series features of the meteorological data sequence segments are obtained, and all local time-series features are fused to obtain high-dimensional time-series dynamic features.

[0009] Preferably, the acquisition of multi-dimensional time-series meteorological data and the combination of a sliding window segmentation method to obtain meteorological data sequence segments include: First, the collected multidimensional time-series meteorological data are preprocessed, including normalization, standardization, and missing value imputation. Then, the preprocessed multidimensional time-series meteorological data is segmented according to the window length and step size to obtain meteorological data sequence segments.

[0010] Preferably, step 2 involves collecting geospatial data and converting it into a fuzzy set, using inference rules to perform fuzzy inference on each geographic attribute in the fuzzy set, and defuzzifying the inference results to obtain geospatial features, including: Convert geospatial data into fuzzy sets based on membership functions; Construct inference rules, use membership functions to determine the activation strength of each inference rule, and obtain the fuzzy inference result based on the activation strength of each inference rule; The geospatial features are obtained by defuzzifying the fuzzy inference results using the centroid method.

[0011] Preferably, step 3, which involves fusing high-dimensional temporal dynamic features and geospatial features using a gating mechanism to obtain high-dimensional spatiotemporal fused features, includes: By combining high-dimensional temporal dynamic features and geospatial features, combined features are obtained; A gating mechanism is used to dynamically adjust the weights of temporal and spatial features in the fusion process to obtain a fused feature vector. The fused feature vector is then subjected to a nonlinear transformation by an activation function to obtain a high-dimensional spatiotemporal fused feature.

[0012] Preferably, step 4, determining the physical residual based on the radiation prediction value and the physical relationship, includes: The time derivative and spatial derivative are calculated based on the radiation prediction values, and the physical residual is determined based on the time derivative, spatial derivative, and physical relationships.

[0013] Preferably, the step of determining the physical residuals based on the radiation prediction values ​​and physical relationships, and combining the physical residuals with the data loss term to obtain the total physical constraint loss function includes: The L2 norm of the physical residuals is used as the physical loss term, and the MSE is used as the data loss term. The learnable weight parameters α and β are dynamically adjusted to perform a weighted sum of the physical loss term and the data loss term, resulting in the total physical constraint loss function.

[0014] Preferably, the time derivative term is:

[0015] Spatial derivative term:

[0016] Physical residuals:

[0017] in, This represents the solar radiation value predicted by the neural network. t Indicates time, while Δ t That is the time step. F The flux of solar radiation, Δ x and Δ y These are the spatial step sizes along the x-axis and y-axis, respectively. Q This represents the source term of solar radiation.

[0018] Preferably, in step 4, the gradient descent method is used to minimize the total loss function of physical constraints to obtain the trained data-driven model.

[0019] Secondly, this application provides a spatiotemporal prediction system for solar radiation based on physical information constraints and neural networks, comprising: The meteorological acquisition module is used to acquire high-dimensional time-series dynamic characteristics of multi-dimensional time-series meteorological data; The geographic data acquisition module is used to collect geospatial data and convert it into a fuzzy set. It uses inference rules to perform fuzzy inference on each geographic attribute in the fuzzy set and defuzzifies the inference results to obtain geospatial features. The fusion module is used to fuse high-dimensional temporal dynamic features and geospatial features using a gating mechanism to obtain high-dimensional spatiotemporal fused features; The prediction module is used to train the data-driven model. The training method is as follows: The high-dimensional spatiotemporal fusion features are input into the data-driven model to obtain the current radiation prediction value. The physical relationship between meteorological data and the current radiation prediction value is determined according to the energy conservation equation. The physical residual is determined according to the radiation prediction value and the physical relationship. The physical residual is combined with the data loss term to obtain the physical constraint total loss function. The parameters of the data-driven model are updated using the physical constraint total loss function and iteratively trained until the physical constraint total loss function is minimized, resulting in a trained data-driven model. Solar radiation values ​​are then predicted based on the trained data-driven model.

[0020] Compared with the prior art, the present invention has the following beneficial technical effects: This application presents a spatiotemporal prediction method for solar radiation based on physical information constraints and neural networks. Through multimodal feature fusion and a physical constraint optimization mechanism, it achieves a dual improvement in the accuracy and reliability of solar radiation prediction. First, by collecting multidimensional time-series meteorological data and extracting high-dimensional dynamic features, combined with fuzzy inference transformation of geospatial data, a gating mechanism is used to achieve deep fusion of spatiotemporal features, effectively solving the prediction bias caused by the fragmentation of spatiotemporal information in traditional methods. More importantly, this method introduces the energy conservation equation into the model training process, constructing physical residuals by calculating the time and spatial derivatives of the predicted values, and combining them with the data loss term to form a physical constraint total loss function. This design ensures that the model not only pursues data fit during optimization but also forces its output results to conform to fundamental thermodynamic laws. Compared to purely data-driven models, this method significantly improves the physical rationality of the prediction results. Especially in extreme weather or data-scarce scenarios, the physical constraint mechanism can effectively correct abnormal predictions caused by data noise, giving the model stronger generalization ability and robustness. Meanwhile, the combination of the gating feature fusion mechanism and the gradient descent optimization algorithm enables accurate representation of high-dimensional spatiotemporal features while ensuring computational efficiency, providing more accurate and reliable technical support for application scenarios such as solar energy resource assessment and photovoltaic power generation prediction.

[0021] This application also proposes a solar radiation spatiotemporal prediction system based on physical information constraints and neural networks, an electronic device, and a computer storage medium, which possess all the advantages of the aforementioned solar radiation spatiotemporal prediction method based on physical information constraints and neural networks. Attached Figure Description

[0022] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a flowchart of the spatiotemporal prediction method for solar radiation based on physical information constraints and neural networks according to the present invention. Figure 2 This is a flowchart of the multi-dimensional temporal dynamic extraction method of the present invention; Figure 3 This is a flowchart of the geospatial nonlinear modeling method based on fuzzy reasoning of the present invention; Figure 4 This is a flowchart of the physical equation consistency constraint method of the present invention. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0025] Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of the application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0026] This application aims to address the following problems with existing solar radiation prediction methods: 1. Problems in refined modeling of spatial geographic information: Insufficient utilization of the geospatial information (latitude, longitude, altitude) of observation points makes it difficult to capture their nonlinear effects on solar radiation and subtle spatial differences.

[0027] 2. Deep extraction of temporal features and spatiotemporal coupling: This addresses the shortcomings of existing methods in effectively extracting complex temporal dynamic features of solar radiation and deeply and organically coupling spatial and temporal information.

[0028] 3. Problem of ensuring physical consistency of prediction results: This addresses the issue that prediction results from purely data-driven models may deviate from basic physical laws, leading to insufficient scientific reliability.

[0029] 4. Improve the accuracy and generalization ability of cross-scale predictions across the country: This addresses the issue that the prediction accuracy and generalization ability of existing models need to be improved when applied to a wide geographical area and different time scales.

[0030] This application presents a spatiotemporal prediction method for solar radiation based on physical information constraints and neural networks. This model combines the Adaptive Neural Fuzzy Inference System (ANFIS), Convolutional Neural Network-Long Short-Term Memory (CNN-LSTM), and Physical Information Neural Network (PINN). ANFIS is used to meticulously model geospatial information (such as latitude, longitude, and altitude), deeply exploring its nonlinear impact on solar radiation distribution and effectively enhancing spatial feature representation. The CNN-LSTM architecture enables deep extraction of complex temporal dynamics and organically integrates with spatial features through a gating mechanism, enhancing the model's adaptability to multidimensional information. Physical equation constraints are introduced into the loss function to guide model training, ensuring the physical authenticity of the prediction results and significantly improving their scientific rigor and credibility. The model demonstrates excellent generalization ability across a wide range of temporal and spatial scales nationwide, making it suitable for various practical scenarios such as solar resource assessment, photovoltaic power plant site selection, and operation and maintenance scheduling, thus possessing high engineering application value.

[0031] A spatiotemporal prediction method for solar radiation based on physical information constraints and neural networks includes the following steps: Step 1: Collect multi-dimensional time-series meteorological data and obtain the high-dimensional time-series dynamic characteristics of the multi-dimensional time-series meteorological data.

[0032] By collecting multidimensional time-series meteorological data (such as temperature, humidity, and wind speed), high-dimensional time-series dynamic features are extracted using time-series analysis techniques. The principle is to capture the dynamic patterns of meteorological elements changing over time, providing key information in the time dimension for subsequent spatiotemporal feature fusion, and ensuring that the model can reflect the time-series dependence of solar radiation.

[0033] Step 2: Collect geospatial data and convert it into a fuzzy set. Use inference rules to perform fuzzy inference on each geographic attribute in the fuzzy set, and defuzzify the inference results to obtain geospatial features.

[0034] Geospatial data (such as topography and latitude / longitude) is transformed into fuzzy sets, and fuzzy inference is performed on geographic attributes through inference rules. Finally, the data is defuzzified to obtain geospatial features. Its function is to transform discrete geographic information into continuous numerical features. The principle is to handle the uncertainty of geographic elements through fuzzy logic, enabling the model to utilize spatial distribution information to improve the geographic adaptability of radiation prediction.

[0035] Step 3: Use a gating mechanism to fuse high-dimensional temporal dynamic features and geospatial features to obtain high-dimensional spatiotemporal fusion features.

[0036] A gating mechanism is employed to fuse high-dimensional temporal dynamic features and geospatial features, generating high-dimensional spatiotemporal fusion features. Its function is to integrate multi-source information across temporal and spatial dimensions. The principle is to dynamically adjust the weights of the two types of features through gating units, thereby addressing the prediction bias caused by the separation of spatiotemporal features in traditional methods and enhancing the model's ability to represent complex spatiotemporal patterns.

[0037] Step 4: Train the data-driven model. The training method is as follows: High-dimensional spatiotemporal fusion features are input into the data to drive the model to obtain the current radiation prediction value. The physical relationship between meteorological data and the current radiation prediction value is determined according to the energy conservation equation. The time derivative and spatial derivative are calculated based on the radiation prediction values. Then, the physical residuals are determined by combining them with the physical relationships. The physical residuals are combined with the data loss terms to obtain the total physical constraint loss function. The parameters of the data-driven model are updated using the total physical constraint loss function. The total physical constraint loss function is minimized using the gradient descent method to obtain the trained data-driven model. Meteorological data is then input into the trained data-driven model to obtain the corresponding radiation prediction values.

[0038] The data-driven model is trained using physical constraints: First, high-dimensional spatiotemporal fusion features are input to obtain radiation prediction values. Then, physical relationships are constructed based on the energy conservation equation. Physical residuals are derived by calculating the temporal / spatial derivatives of the prediction values ​​and combining them with the physical relationships. These residuals are then combined with the data loss term to form the total physical constraint loss function. Finally, the model parameters are optimized using gradient descent. Its purpose is to ensure that the prediction results conform to physical laws. The principle is to embed physical knowledge into the data-driven framework, solving the problem that pure data models may violate physical laws, and significantly improving the reliability and generalization of predictions.

[0039] Example 1 A spatiotemporal prediction method for solar radiation based on physical information constraints and neural networks includes the following steps: Step 1: Use a multi-dimensional time-series dynamic extraction module to collect multi-dimensional time-series meteorological data and obtain high-dimensional time-series dynamic features of multi-source time-series meteorological data.

[0040] The multi-dimensional time-series dynamic extraction module extracts dynamic features from multi-dimensional time-series meteorological data, laying a solid foundation for subsequent spatial modeling and spatiotemporal fusion. This module mainly collects multi-dimensional meteorological variables such as solar radiation, temperature, humidity, and visibility. These variables are usually input in the form of time series, reflecting the dynamic changes and complex correlations of the meteorological environment.

[0041] The multi-dimensional time-series dynamic extraction module includes a data preprocessing unit, a sliding window segmentation unit, and a feature extraction unit.

[0042] 1. Data Preprocessing Unit The data preprocessing unit performs normalization and standardization on the raw meteorological data to ensure that meteorological data with different dimensions and distributions can be modeled at the same scale. In addition, to address the common problem of missing values ​​in actual observation data, the preprocessing unit also uses interpolation, mean imputation and other methods to complete the data, improving the completeness of the data and the robustness of subsequent modeling. After preprocessing, the multidimensional meteorological time series data is input into the sliding window segmentation unit.

[0043] 2. Sliding window splitting unit The sliding window segmentation unit divides the long-term series into multiple meteorological data sequence segments according to the set window length and step size. Each meteorological data sequence segment represents a local time segment. This segmentation method helps to capture the local dynamic features of meteorological variables in the short term, such as changing trends, abrupt changes, and periodicity. It also provides more granular input for subsequent feature extraction.

[0044] 3. Feature Extraction Unit The feature extraction unit performs local feature extraction on each meteorological data sequence segment. Through the sliding operation of multiple convolutional kernels, it can automatically learn and extract important dynamic patterns in local time segments, such as sudden changes and periodic fluctuations, thereby obtaining local time-series features. The convolution operation not only improves the efficiency of feature extraction, but also effectively reduces the impact of noise on modeling, allowing the model to focus more on key time-series signals.

[0045] The convolutional local temporal features are input into a multi-layer LSTM temporal modeling unit. LSTM (Long Short-Term Memory) employs a gating mechanism, effectively capturing long-term dependencies and complex dynamic changes in time series data. Through multi-layer stacking, LSTM can perform deep modeling of temporal features at different levels and scales, obtaining high-dimensional dynamic temporal features that preserve local dynamic information while incorporating global temporal trends. This process significantly enhances the model's ability to express meteorological time series data.

[0046] The high-dimensional temporal dynamic features output by the multi-dimensional temporal dynamic extraction module will serve as a key input to the subsequent spatiotemporal feature fusion module. These high-dimensional temporal dynamic features not only contain short-term and long-term dynamic information of meteorological variables but also provide a rich foundation for the model to understand the evolution of meteorological systems. Through the synergistic effect of core components such as data preprocessing, sliding window segmentation, convolutional feature extraction, and LSTM deep modeling, the multi-dimensional temporal dynamic extraction module achieves efficient and comprehensive representation of multi-dimensional meteorological time-series data.

[0047] See Figure 2 The following describes the working process of the multi-dimensional time series dynamic extraction module, including the following steps: S1.1 Receives multidimensional time-series meteorological data collected from external data sources (such as weather stations or remote sensing platforms), including solar radiation, temperature, humidity, visibility, etc. The data preprocessing unit normalizes, standardizes, and imputes missing values ​​in the multidimensional time-series meteorological data to ensure the stability and consistency of subsequent processing.

[0048] S1.2. The preprocessed multidimensional time-series meteorological data is fed into a sliding window segmentation unit. This unit divides the long-term series into multiple overlapping or non-overlapping meteorological data sequence segments according to the set window length and step size. Each meteorological data sequence segment represents a local time segment. The meteorological data sequence segments are then input into a one-dimensional convolutional (Conv1d) feature extraction unit. Multiple convolutional kernels extract local features for each time segment, outputting a multi-channel local time-series feature matrix. Convolutional operations can effectively capture short-term trends, abrupt changes, and periodicity. The convolved local time-series features are then fed into a multi-layer LSTM time-series modeling unit, which uses a recurrent neural network structure.

[0049] The sliding window mechanism slides the input sequence X = [x1, x2, ..., xT] with a window size of k to obtain the subsequence Xt = [xt, xt+1, ..., xt+k-1]. For each window, a one-dimensional convolution operation is performed using a convolution kernel w, calculated as follows:

[0050] Where ht is the convolution output and b is the bias.

[0051] S1.3, Modeling long-term dependencies and complex dynamic changes in the sequence. Stacking multiple LSTMs enables the model to capture temporal features from different levels, with the LSTM output serving as high-dimensional temporal dynamic features.

[0052] This high-dimensional temporal dynamic feature can be directly used in subsequent spatiotemporal feature fusion modules, or it can interact with other modules (such as the physical constraint module). Throughout the process, data flows sequentially between units, and the output of each step provides a rich feature foundation for the next step, realizing the layer-by-layer refinement and information enhancement from raw data to high-dimensional temporal features.

[0053] The LSTM cell is recursively calculated as follows:

[0054]

[0055]

[0056]

[0057]

[0058] By stacking multiple layers, LSTM can enhance the ability to model complex temporal dynamics.

[0059] Step 2: Collect geospatial data and convert it into a fuzzy set. Use inference rules to perform fuzzy inference on each geographic attribute in the fuzzy set, and defuzzify the inference results to obtain geospatial features.

[0060] In this embodiment, a geospatial nonlinear modeling module based on fuzzy inference is used to obtain geospatial features.

[0061] The fuzzy inference-based geospatial nonlinear modeling module performs nonlinear modeling of geospatial information, fully exploring the complex influence of geographical attributes such as longitude, latitude, and altitude on target variables (e.g., solar radiation). Unlike traditional linear modeling methods, it employs fuzzy inference to effectively express and utilize the continuity and uncertainty of geospatial data, thereby enhancing the model's ability to characterize spatial heterogeneity.

[0062] The geospatial nonlinear modeling module for fuzzy inference includes a geographic data fuzzification unit, a membership function design and parameter adjustment unit, a fuzzy inference rule base construction unit, and a fuzzy inference engine unit.

[0063] 1. Geographic data fuzzification units transform raw geospatial data (such as longitude, latitude, and altitude) into fuzzy sets. By designing appropriate membership functions, each geographic attribute is divided into several fuzzy categories (such as "high," "medium," and "low"), and a membership degree is assigned to each category. This fuzzification process can better reflect the gradual changes and fuzziness of geographic attributes in the actual environment, avoiding information loss caused by hard segmentation.

[0064] 2. Membership Function Design and Parameter Adjustment Unit: This unit selects an appropriate membership function type for each geographic attribute and adjusts the parameters based on the actual data distribution. The rational design of the membership function directly affects the scientific validity and effectiveness of the fuzzification results; therefore, this unit typically requires optimization by combining domain knowledge and data analysis results.

[0065] 3. The fuzzy inference rule base construction unit establishes a series of "if-then" type fuzzy inference rules through expert knowledge, historical data analysis, or automated rule mining. For example, a rule such as "if the altitude is high and the latitude is low, then the radiation is strong" can be set. The richness and rationality of the rule base directly determine the model's ability to express nonlinear relationships in geospatial data.

[0066] 4. The fuzzy inference engine unit calculates the activation degree of each rule based on the membership degree of the input geographic attributes and the rule base, and synthesizes the outputs of all rules to obtain a fuzzy inference result. This process can flexibly integrate the influence of multiple geographic attributes, reflecting the complex interaction relationships between spatial factors.

[0067] 5. Defuzzification unit: This unit transforms the fuzzy results output by the fuzzy inference engine into specific numerical spatial features. Common methods include the centroid method. This step ensures that the final output spatial features can be seamlessly integrated with other modules (temporal features, spatiotemporal fusion module) to participate in subsequent modeling and prediction.

[0068] This fuzzy inference-based geospatial nonlinear modeling module achieves nonlinear and flexible modeling of geospatial information through the collaborative work of core components such as geographic data fuzzification, membership function design, rule base construction, fuzzy inference, and defuzzification. Its functional positioning is not only to supplement spatial information with temporal features, but also to enhance the model's ability to perceive and express spatial heterogeneity and complex geographical relationships, providing a solid spatial foundation for subsequent spatiotemporal feature fusion and physical constraints.

[0069] See Figure 3 The following describes the working process of the geospatial nonlinear modeling module for fuzzy inference, including the following steps: S2.1 Collect geospatial data. The geographic data preprocessing unit normalizes and standardizes the input data to ensure that data of different dimensions can be processed uniformly.

[0070] S2.2 The preprocessed geographic data is fed into the fuzzification unit, which uses a preset membership function to transform continuous geographic variables into fuzzy sets and outputs the membership degree of each attribute under different fuzzy sets.

[0071] Fuzzification: For each input variable The system defines a set of membership functions. This is used to characterize the degree to which a function belongs to a certain fuzzy set (such as "high", "medium", "low"). Taking the "high" fuzzy set at altitude h as an example, its sigmoid membership function can be defined as:

[0072] in, a , b These are learnable parameters that control the steepness and center position of the curve.

[0073] S2.3 The fuzzy inference engine combines the membership degree of the input with a pre-defined "if-then" rule base to perform fuzzy inference on geographical attributes. The inference engine will calculate the fuzzy output based on the activation degree of each rule.

[0074] Fuzzy rule matching and activation: The system presets M fuzzy rules, in the following form:

[0075] in: For the fuzzy subset corresponding to the 𝑘th input in the 𝑗th rule; For consequent functions, linear models are often used.

[0076] For each rule, calculate its activation strength (matching degree) based on the membership degree of the input variables:

[0077] Rule output synthesis: based on the activation strength of each rule. and corresponding output Calculate the weighted average of the overall fuzzy output:

[0078] S2.4 The fuzzification unit transforms the fuzzy output into specific numerical spatial features using methods such as the centroid method.

[0079] The resulting geospatial features can not only serve as independent inputs for subsequent spatiotemporal feature fusion, but also interact with modules such as temporal features and physical constraints. The entire process achieves nonlinear mapping and feature enhancement of geospatial information, providing a spatial dimension to the model.

[0080] Deblurring: The system uses the centroid method to clarify the fuzzy output, and finally outputs a numerical geospatial feature y, which can be directly used for deep feature fusion with a time series model (CNN-LSTM).

[0081] Finally, by inputting geospatial variables such as longitude (lon), latitude (lat), and altitude (h), the output is a geospatial feature vector after fuzzy inference. Each dimension represents a comprehensive expression of a specific spatial dimension in a fuzzy rule space.

[0082] Step 3: Fuse the high-dimensional temporal dynamic features and geospatial features to obtain high-dimensional spatiotemporal fusion features.

[0083] In this embodiment, a gated spatiotemporal feature fusion module is used to obtain high-dimensional spatiotemporal fusion features. The gated spatiotemporal feature fusion module realizes efficient fusion of temporal dynamic features and spatial features, thereby fully exploring the intrinsic relationship between meteorological spatiotemporal data and improving the model's comprehensive expressive ability and prediction performance.

[0084] The gated spatiotemporal feature fusion module includes a feature splicing / weighting unit, a gating mechanism unit, an activation function unit, and a feature output unit, forming a flexible and adaptive spatiotemporal feature fusion mechanism.

[0085] 1. Feature concatenation / weighting unit: This unit initially integrates high-dimensional temporal dynamic features and geospatial features. Common approaches include directly concatenating the two types of feature vectors or combining them through weighted summation. Concatenation preserves the complete information of each feature, while weighting can highlight the influence of a particular type of feature according to actual needs.

[0086] 2. Gating Mechanism Unit: Gating mechanisms (such as gated weighting, GRU gating, attention mechanisms, etc.) dynamically adjust the weight distribution of temporal and spatial features during the fusion process based on the current input features, thereby obtaining a fused feature vector. For example, a gated weighting mechanism assigns a weight coefficient to each type of feature. The magnitude of the weight is adaptively generated by passing the input features through a neural network layer (usually a fully connected layer with sigmoid activation). Specifically, the calculation process of the gating mechanism can be represented as follows:

[0087]

[0088] Where z is the gating weight, ⊙ represents element-wise multiplication, and σ is the sigmoid activation function. and These are temporal and spatial features, respectively. Through a gating mechanism, the fusion ratio of temporal and spatial features can be flexibly adjusted according to different input contexts, achieving organic integration of information.

[0089] 3. Activation function units, such as sigmoid, tanh, or ReLU. The introduction of activation functions enhances the nonlinear expressive power of the model, making the fused feature vectors more discriminative and adaptable in high-dimensional space, thus obtaining high-dimensional spatiotemporal fusion features. Furthermore, activation functions can prevent gradient vanishing or exploding, improving the stability of model training.

[0090] 4. Feature Output Unit: This unit outputs the fused high-dimensional spatiotemporal features, providing input for the subsequent physical consistency constraint module and the final prediction module. The high-dimensional spatiotemporal fusion features not only contain the temporal dynamics of meteorological data but also fully reflect the nonlinear effects of geographic space.

[0091] This gated spatiotemporal feature fusion module achieves deep fusion of spatiotemporal features through the synergistic effect of core components such as feature splicing / weighting, gating mechanisms, and activation functions. Its function is not only to enhance the model's ability to model complex meteorological phenomena, but also to provide high-quality, physically reasonable feature inputs for subsequent physical constraints and predictions, making it a key link in the entire system to achieve high-precision, strong generalization predictions.

[0092] See Figure 4 The working principle of the gated spatiotemporal feature fusion module is explained, including the following steps: S3.1 The feature splicing / weighting unit initially combines high-dimensional temporal dynamic features and geospatial features to form a feature vector containing spatiotemporal information; S3.2 The gating mechanism unit (such as gated weighting, GRU gating, etc.) dynamically calculates the fusion weights of temporal and spatial features based on the current feature vector to obtain the fused feature vector. The gating mechanism can adaptively adjust the contribution ratio of the two types of features to achieve organic fusion of information.

[0093] The gating mechanism first calculates the gating weight z:

[0094] Then, the fused output is obtained through a weighted method:

[0095] Where ⊙ represents element-wise multiplication, and σ is the sigmoid activation.

[0096] S3.3 The fused feature vector undergoes a nonlinear transformation using activation function units (such as sigmoid, tanh, etc.) to further enhance the expressive power of the features. Finally, the output high-dimensional spatiotemporal fused features can be used as input for the final prediction or interact with the physical constraint module below to participate in loss calculation and model optimization.

[0097] Nonlinear transformation of fused features:

[0098] Step 4: Train the data-driven model. The training method is as follows: By inputting multidimensional meteorological data sequences and high-dimensional spatiotemporal fusion features into the data-driven model, the current radiation prediction value is obtained, and the physical relationship between meteorological data and radiation prediction value is determined according to the energy conservation equation. The time derivative and spatial derivative are calculated based on the radiation prediction values. Then, the physical residuals are determined in combination with the physical relationships. The physical residuals are used as the physical loss term, and the MSE is used as the data loss term. The physical loss term and the data loss term are weighted and summed to obtain the total physical constraint loss function. The parameters of the data-driven model are updated using the total physical constraint loss function. The total physical constraint loss function is minimized using the gradient descent method to obtain the trained data-driven model. Meteorological data is then input into the trained data-driven model to obtain the corresponding radiation prediction values.

[0099] In this embodiment, a physical equation consistency constraint module is used to train the data-driven model, deeply integrating physics knowledge with the data-driven model to ensure that the model's predictions not only closely approximate the observed data numerically but also possess rationality in terms of physical mechanisms. Its main function is to improve the model's generalization ability and scientific interpretability by introducing physical equations to impose consistency constraints on the model output.

[0100] The physical equation consistency constraint module includes a physical equation modeling unit, a physical residual calculation unit, a loss function integration unit, and a backpropagation and optimization unit.

[0101] 1. The physical equation modeling unit uses the energy conservation equation as the basic physical constraint. It combines meteorological input data (temperature, humidity, visibility, etc.) with the current predicted values ​​of the data-driven model, substitutes them into the physical equation, and calculates the physical relationship that should be satisfied in theory, that is, the relationship between meteorological data and radiation values.

[0102] 2. The physical residual calculation unit quantifies the deviation between the current predicted value output by the data-driven model and the actual physical relationship, thus determining the physical residual of the data-driven model. Specifically, it calculates the time derivative of the radiation prediction value using the second-order central difference method, calculates the spatial derivative using the five-point difference scheme, and compares it with the theoretical value on the right-hand side of the physical equation to obtain the physical residual term.

[0103] For example, the time derivative term can be calculated using the finite difference method, and the spatial derivative term can be calculated using the central difference scheme.

[0104] 3. The loss function integration unit uses physical residuals as the physical loss term and MSE as the data loss term. It weights and fuses the physical residuals and data function to form the total physical constraint loss function. Specifically, MSE is used as the data loss term, and the L2 norm of the physical residuals is used as the physical loss term. These are dynamically adjusted using learnable weight parameters α and β. This mechanism encourages the model to actively avoid physically unreasonable solutions while pursuing prediction accuracy.

[0105] Backpropagation and Optimization Unit: This unit updates the parameters of the entire model using the total physical constraint loss function. It employs the Adam optimizer with a learning rate of 1e-4, minimizing the total physical constraint loss function through gradient descent. Due to the introduction of the physical loss term, the model continuously adjusts its parameters during training, gradually bringing the output closer to an ideal state that satisfies both the observed data and the physical equation constraints.

[0106] The working principle of the physical equation consistency constraint module is explained in detail below, including the following steps: S4.1 Input the high-dimensional spatiotemporal fusion features into the data-driven model to obtain the current predicted value. Substitute the meteorological data and the current radiation prediction value into the energy conservation equation to calculate the physical relationship between the meteorological data and the radiation value that should theoretically be satisfied. That is, use the energy conservation equation to perform a physical consistency check on the prediction results.

[0107] The energy conservation equation is:

[0108] Where S is the radiation quantity, F is the flux, and Q is the source term.

[0109] S4.2 Calculate the time derivative and spatial derivative based on the radiation prediction value, and then determine the physical residual of the current radiation prediction value by combining the physical relationship. Use the finite difference method to calculate the residual between the current radiation prediction value and the physical law to obtain the physical residual value.

[0110] Model output The residuals from the physical equations are calculated using the finite difference method: Time derivative term:

[0111] Spatial derivative term:

[0112] Physical residuals:

[0113] in, This represents the solar radiation value predicted by the neural network; it is the amount of solar radiation at a specific time and location. t Indicates time, while Δ t This refers to the time step, which is the time interval between adjacent time points. In describing the spatial distribution and transmission of solar radiation, F Refers to the flux of solar radiation, a vector that quantifies the amount of solar radiation energy passing through a unit area per unit time. Spatial grid points pass through... i and j Indexed, representing discrete positions on the x-axis (longitude) and y-axis (latitude), respectively. Correspondingly, Δ x and Δ y These are the spatial step sizes along the x-axis and y-axis, respectively, representing the spatial distance between adjacent grid points. Finally, QThis indicates the source term of solar radiation, which in physics usually refers to the external input or consumption of solar radiation. For example, the absorption and scattering of solar radiation by the atmosphere can lead to an increase or decrease in local solar radiation.

[0114] S4.3. The physical residuals and MSE loss are weighted and combined to form the total physical constraint loss function. This total physical constraint loss function not only reflects the error between the model prediction and the actual observation, but also reflects the consistency between the radiation prediction output by the model and the physical laws.

[0115] Total loss function due to physical constraints:

[0116] Here, α and β are learnable weight parameters that are updated by the Adam optimizer.

[0117] S4.4 The backpropagation and optimization unit uses the Adam optimizer to optimize the model parameters, ensuring that the model follows basic physical laws while improving prediction accuracy, and finally obtains the trained data-driven model, namely the radiation prediction model.

[0118] The physical equation consistency constraint module works closely with the physical constraint module, the spatiotemporal feature fusion module, and the prediction output module. Through the feedback mechanism of the loss function, it achieves an organic combination of data-driven and physics-driven approaches, thereby improving the model's generalization ability and physical rationality.

[0119] This embodiment provides a spatiotemporal prediction method for solar radiation based on physical information constraints and neural networks, which significantly improves the prediction accuracy, robustness, and applicability of the model in complex spatial environments, and has the following advantages: (1) Significantly improves prediction accuracy and generalization ability: Significantly Enhanced Spatial Intelligent Modeling Capabilities: This invention pioneered the introduction of an Adaptive Network Fuzzy Inference System (ANFIS) to model the spatial influencing factors of solar radiation, particularly for geographical variables with significant nonlinear characteristics such as latitude, longitude, and altitude. Through fuzzy rule inference and parameter self-learning mechanisms, it achieves deeper modeling of geospatial influences. This method overcomes the limitations of traditional models that merely treat geographical information as static additional features or perform coarse spatial interpolation, enabling the model to adaptively learn the pattern differences in solar radiation response across different regions, forming region-specific predictive "genes." For example, in regions with significantly different topography, such as plateaus, coastlines, and deserts, the ANFIS module can identify the systematic influence of implicit features such as altitude, humidity, and topographic relief on radiation distribution and produce differentiated nonlinear responses. This capability significantly enhances the model's spatial generalization performance, providing strong support for wide-area deployment nationwide.

[0120] Construction of a Deep Spatiotemporal Feature Coupling Structure: The CNN-LSTM hybrid neural network structure constructed in this invention achieves efficient modeling of the temporal evolution characteristics of solar radiation through collaborative learning. Specifically, the Convolutional Neural Network (CNN) is used to mine local features such as local temporal structure, diurnal variation patterns, and anomalous peaks in solar radiation data, while the Long Short-Term Memory Network (LSTM) captures long-term dependent information such as long-term trends, periodic changes, and seasonal variations. Unlike traditional methods, this model adopts a gated fusion mechanism, dynamically weighting and integrating the spatial semantic information extracted by ANFIS with the temporal information extracted by CNN-LSTM. This allows the model to adaptively adjust the degree of attention to spatial or temporal factors based on the features of the current input data. Compared with simple feature concatenation methods, this fusion strategy is more adaptable and robust in large-scale spatiotemporal prediction tasks. Comprehensive evaluation results show that the RMSE and MAE indicators of this model on typical regional datasets are improved by 5% to 15% compared with models using LSTM, CNN-LSTM alone, or models without integrated spatial modeling modules, especially in complex terrain and climate transition zones.

[0121] (2) Enhance the physical consistency and scientific reliability of prediction results: This invention integrates the Physics-Informed Neural Network (PINN) concept into its model design, incorporating the fundamental physical laws of solar radiation into the model training process, thus forming a dual guidance mechanism of data-driven and physical constraints. Specifically, the model introduces soft constraints such as clear-sky radiation models, energy conservation equations, diurnal variation envelope constraints, and the boundary of the maximum theoretical value of radiation, and adds corresponding physical residual regularization terms to the loss function. These physical constraints not only effectively prevent the model from producing "illusionary predictions" that defy common sense at night, in extreme weather, or in areas with sparse training data, but also endow the model with stronger extrapolation and interpretability to a certain extent. For example, in mountainous or desert areas lacking observational data, the model can still maintain a reasonable radiation distribution pattern without predicting unreliable values ​​beyond the physical range, thereby greatly enhancing its reliability in wide-area deployment and long-term time-series prediction.

[0122] (3) Enhanced model interpretability and debugging capabilities: Beyond accuracy and physicality, this invention also features targeted design improvements in model interpretability, making it more suitable for understanding and redevelopment in scientific research and industrial scenarios. First, the fuzzy rule base and membership functions of the ANFIS module possess a transparent semantic structure, allowing users to observe how the model categorizes geographical information into different fuzzy intervals and determine its influence direction and weight through "if-then" rules. Second, the gating signal weights in the gating fusion mechanism can be used to analyze the model's tendency to allocate weights to spatial and temporal features at different times and in different regions, further understanding the model's internal decision-making logic. Third, the changing trajectories of various physical residual terms in the PINN structure can also serve as a diagnostic tool, helping researchers locate the position and degree of deviation from physical laws during model training, thus enabling fine-tuning or local correction. These mechanisms not only enhance the model's visualization and controllability but also provide a basis for subsequent tasks such as domain transfer learning and model fine-tuning.

[0123] (4) Wider range of applicable scenarios and application prospects: This invention offers high flexibility and adaptability in model structure design and data fusion methods, making it particularly suitable for nationwide solar radiation prediction scenarios characterized by diverse climates and complex terrains. By effectively modeling spatial heterogeneity, the model maintains good predictive performance across regions with vastly different meteorological conditions, such as the southeast coast, the North China Plain, the Southwest Plateau, and inland basins, achieving truly unified national model deployment. Furthermore, since CNN-LSTM inherently supports multi-timescale learning, this method can be flexibly applied to minute-level, hourly, and even daily solar radiation prediction tasks without modifying the core structure, thus meeting the needs of various scenarios such as weather warnings, photovoltaic power generation scheduling, and agricultural and forestry ecological simulation. More importantly, thanks to the embedding of physical laws in PINN, the model maintains strong anti-interference capabilities even in scenarios with missing or fluctuating observational data, making it suitable for tasks such as remote sensing observation in sparse areas and future scenario simulation, providing technical support for solar energy development and utilization in low-resource areas.

[0124] In summary, this application presents a novel spatiotemporal prediction method for solar radiation that integrates neural networks, fuzzy systems, and physical models. This method significantly overcomes multiple bottlenecks in existing methods regarding spatial modeling, temporal coupling, physical consistency, and model interpretability. Both experimental results and theoretical analysis demonstrate that this method exhibits significant advantages in prediction accuracy, generalization ability, physical reliability, model transparency, and deployment breadth. Its multi-layered structure and modular design also give it excellent scalability and engineering application potential, providing a solid foundation for promoting the intelligent utilization of solar energy resources, the construction of green energy systems, and scientific research on renewable energy.

[0125] Example 2 Correspondingly, this application also provides a solar radiation spatiotemporal prediction system based on physical information constraints and neural networks, including: The meteorological acquisition module is used to acquire high-dimensional time-series dynamic characteristics of multi-dimensional time-series meteorological data; The geographic data acquisition module is used to collect geospatial data and convert it into a fuzzy set. It uses inference rules to perform fuzzy inference on each geographic attribute in the fuzzy set and defuzzifies the inference results to obtain geospatial features. The fusion module is used to fuse high-dimensional temporal dynamic features and geospatial features using a gating mechanism to obtain high-dimensional spatiotemporal fused features; The prediction module is used to train the data-driven model. The training method is as follows: The high-dimensional spatiotemporal fusion features are input into the data-driven model to obtain the current radiation prediction value. The physical relationship between meteorological data and the current radiation prediction value is determined according to the energy conservation equation. The physical residual is determined according to the radiation prediction value and the physical relationship. The physical residual is combined with the data loss term to obtain the physical constraint total loss function. The parameters of the data-driven model are updated using the physical constraint total loss function and iteratively trained until the physical constraint total loss function is minimized, resulting in a trained data-driven model. Solar radiation values ​​are then predicted based on the trained data-driven model.

[0126] It should be noted that, in the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules may be combined or integrated into another device, or some features may be ignored or not executed. The modules described as separate components may or may not be physically separated. The components shown as modules may be one or more physical units, that is, they may be located in one place or distributed in multiple different places. Some or all of the modules can be selected to achieve the purpose of the solution in this embodiment according to actual needs.

[0127] Furthermore, in the various embodiments of the present invention, the modules can be integrated into one processing unit, or each module can exist physically separately, or two or more modules can be integrated into one unit. The integrated unit described above can be implemented in hardware or as a software functional unit.

[0128] An electronic device provided in this application includes a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it implements the steps of the solar radiation spatiotemporal prediction method based on physical information constraints and neural networks as described in any of the above embodiments.

[0129] Another electronic device provided in this application embodiment may further include: an input port connected to a processor for transmitting multimodal data collected by an external acquisition device to the processor; a display unit connected to the processor for displaying the processor's processing results to the outside world; and a communication module connected to the processor for enabling communication between the electronic device and the outside world. The display unit may be a display panel, a laser scanning display, etc.; the communication method adopted by the communication module includes, but is not limited to, Mobile High Definition Link (HML), Universal Serial Bus (USB), High Definition Multimedia Interface (HDMI), and wireless connection (including Wi-Fi, Bluetooth, Bluetooth Low Energy, and IEEE 802.11s-based communication technology).

[0130] This application provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it implements the steps of the solar radiation spatiotemporal prediction method based on physical information constraints and neural networks as described in any of the above embodiments.

[0131] For descriptions of relevant parts of the solar radiation spatiotemporal prediction system, electronic device, and computer-readable storage medium based on physical information constraints and neural networks provided in this application, please refer to the detailed descriptions of the corresponding parts in the solar radiation spatiotemporal prediction method based on physical information constraints and neural networks provided in this application, which will not be repeated here. Furthermore, parts of the technical solutions provided in this application that are consistent with the implementation principles of corresponding technical solutions in the prior art have not been described in detail to avoid excessive elaboration.

[0132] The above content is only for illustrating the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. Any modifications made to the technical solution based on the technical concept proposed in this invention shall fall within the scope of protection of the claims of this invention.

Claims

1. A spatiotemporal prediction method for solar radiation based on physical information constraints and neural networks, characterized in that, Includes the following steps: Step 1: Obtain the high-dimensional time-series dynamic characteristics of multi-dimensional time-series meteorological data; Step 2: Collect geospatial data and convert it into a fuzzy set. Use inference rules to perform fuzzy inference on each geographic attribute in the fuzzy set, and defuzzify the inference results to obtain geospatial features. Step 3: Use a gating mechanism to fuse high-dimensional temporal dynamic features and geospatial features to obtain high-dimensional spatiotemporal fusion features; Step 4: Train the data-driven model. The training method is as follows: The high-dimensional spatiotemporal fusion features are input into the data-driven model to obtain the current radiation prediction value. The physical relationship between meteorological data and the current radiation prediction value is determined according to the energy conservation equation. The physical residual is determined according to the radiation prediction value and the physical relationship. The physical residual is combined with the data loss term to obtain the physical constraint total loss function. The parameters of the data-driven model are updated using the physical constraint total loss function and iteratively trained until the physical constraint total loss function is minimized, resulting in a trained data-driven model. Solar radiation values ​​are then predicted based on the trained data-driven model.

2. The spatiotemporal prediction method for solar radiation based on physical information constraints and neural networks according to claim 1, characterized in that, Step 1 describes the use of a multi-dimensional time-series dynamic extraction module to collect multi-dimensional time-series meteorological data and obtain high-dimensional time-series dynamic features of the multi-source time-series meteorological data, including: Multidimensional time-series meteorological data are collected and combined with the sliding window segmentation method to divide the multidimensional time-series meteorological data into meteorological data sequence segments. Local time-series features of the meteorological data sequence segments are obtained, and all local time-series features are fused to obtain high-dimensional time-series dynamic features.

3. The spatiotemporal prediction method for solar radiation based on physical information constraints and neural networks according to claim 2, characterized in that, The process of collecting multidimensional time-series meteorological data and combining it with a sliding window segmentation method to obtain meteorological data sequence segments includes: First, the collected multidimensional time-series meteorological data are preprocessed, including normalization, standardization, and missing value imputation. Then, the preprocessed multidimensional time-series meteorological data is segmented according to the window length and step size to obtain meteorological data sequence segments.

4. The spatiotemporal prediction method for solar radiation based on physical information constraints and neural networks according to claim 1, characterized in that, Step 2 involves collecting geospatial data and converting it into a fuzzy set. Fuzzy inference rules are then applied to each geographic attribute within the fuzzy set. The inference results are defuzzified to obtain the geospatial features, including: Convert geospatial data into fuzzy sets based on membership functions; Construct inference rules, use membership functions to determine the activation strength of each inference rule, and obtain the fuzzy inference result based on the activation strength of each inference rule; The geospatial features are obtained by defuzzifying the fuzzy inference results using the centroid method.

5. The spatiotemporal prediction method for solar radiation based on physical information constraints and neural networks according to claim 1, characterized in that, Step 3 describes the use of a gating mechanism to fuse high-dimensional temporal dynamic features and geospatial features, resulting in high-dimensional spatiotemporal fusion features, including: By combining high-dimensional temporal dynamic features and geospatial features, combined features are obtained; A gating mechanism is used to dynamically adjust the weights of temporal and spatial features in the fusion process to obtain a fused feature vector. The fused feature vector is then subjected to a nonlinear transformation by an activation function to obtain a high-dimensional spatiotemporal fused feature.

6. The spatiotemporal prediction method for solar radiation based on physical information constraints and neural networks according to claim 1, characterized in that, Step 4, determining the physical residual based on the radiation prediction value and the physical relationship, includes: The time derivative and spatial derivative are calculated based on the radiation prediction values, and the physical residual is determined based on the time derivative, spatial derivative, and physical relationships.

7. A method for spatiotemporal prediction of solar radiation based on physical information constraints and neural networks according to claim 1 or 6, characterized in that, The process of determining the physical residuals based on radiation predictions and physical relationships, and then combining the physical residuals with data loss terms to obtain the total physical constraint loss function includes: The L2 norm of the physical residuals is used as the physical loss term, and the MSE is used as the data loss term. The learnable weight parameters α and β are dynamically adjusted to perform a weighted sum of the physical loss term and the data loss term, resulting in the total physical constraint loss function.

8. The spatiotemporal prediction method for solar radiation based on physical information constraints and neural networks according to claim 7, characterized in that, The time derivative term: Spatial derivative term: Physical residuals: in, This represents the solar radiation value predicted by the neural network. t Indicates time, while Δ t That is the time step. F The flux of solar radiation, Δ x and Δ y These are the spatial step sizes along the x-axis and y-axis, respectively. Q This represents the source term of solar radiation.

9. The spatiotemporal prediction method for solar radiation based on physical information constraints and neural networks according to claim 7, characterized in that, In step 4, the gradient descent method is used to minimize the total loss function of physical constraints, resulting in the trained data-driven model.

10. A spatiotemporal prediction system for solar radiation based on physical information constraints and neural networks, characterized in that, include: The meteorological acquisition module is used to acquire high-dimensional time-series dynamic characteristics of multi-dimensional time-series meteorological data; The geographic data acquisition module is used to collect geospatial data and convert it into a fuzzy set. It uses inference rules to perform fuzzy inference on each geographic attribute in the fuzzy set and defuzzifies the inference results to obtain geospatial features. The fusion module is used to fuse high-dimensional temporal dynamic features and geospatial features using a gating mechanism to obtain high-dimensional spatiotemporal fused features; The prediction module is used to train the data-driven model. The training method is as follows: The high-dimensional spatiotemporal fusion features are input into the data-driven model to obtain the current radiation prediction value. The physical relationship between meteorological data and the current radiation prediction value is determined according to the energy conservation equation. The physical residual is determined according to the radiation prediction value and the physical relationship. The physical residual is combined with the data loss term to obtain the physical constraint total loss function. The parameters of the data-driven model are updated using the physical constraint total loss function and iteratively trained until the physical constraint total loss function is minimized, resulting in a trained data-driven model. Solar radiation values ​​are then predicted based on the trained data-driven model.

Images