Multiscale intelligent photovoltaic power prediction method and system based on multiple spatiotemporal meteorological elements

Abstract

The application discloses a kind of photovoltaic power multiscale intelligent prediction method and system based on multivariate space-time meteorological elements, belong to new energy power prediction technical field.The method includes: based on Hausdorff dimension adaptive coding realizes multiscale fractal space-time representation;Wavelet neural operator is used in wavelet domain to learn function space mapping to reduce computational complexity;Lyapunov exponent guided chaos attention mechanism is introduced to enhance extreme event modeling;Physical information neural network embedding energy conservation and other physical constraints are constructed;Dynamic causal graph neural network modeling time-varying interaction between meteorological elements is established;Hierarchical Bayesian ensemble is used to integrate multiscale prediction results and quantify uncertainty.The application realizes full-scale accurate prediction from second level to seasonal level, significantly improves the prediction robustness, computational efficiency and physical consistency under extreme weather.

Description

Technical Field

[0001] This invention belongs to the field of new energy power prediction technology, and in particular relates to a multi-scale intelligent prediction method and system for photovoltaic power based on multiple spatiotemporal meteorological elements. Background Technology

[0002] In existing technologies, photovoltaic power prediction mainly relies on various prediction schemes based on statistical methods, machine learning models, and deep learning frameworks. Traditional statistical methods, such as autoregressive integral moving average models, can effectively model stationary time-series data, featuring high computational efficiency and strong model interpretability. With the development of artificial intelligence technology, machine learning methods such as support vector machines and random forests have been introduced into photovoltaic power prediction, achieving preliminary fitting of nonlinear meteorological relationships through feature engineering and model training. In recent years, deep learning models such as convolutional neural networks, long short-term memory networks, and Transformers have been gradually applied to this field, capable of extracting spatiotemporal features from historical meteorological and power data, and achieving high prediction accuracy under certain conditions. These methods provide fundamental technical support for short-term power forecasting and scheduling of photovoltaic power plants in practical engineering, especially performing well in scenarios with high data quality and stable meteorological conditions.

[0003] However, existing photovoltaic power prediction methods still have the following limitations: First, traditional methods usually process prediction tasks at different time scales independently, failing to effectively model the coupling effect between scales, making it difficult to coordinate and optimize multi-scale prediction results; Second, convolutional neural networks, due to their limited local receptive field, struggle to capture long-distance spatiotemporal dependencies, while global attention mechanisms such as Transformer, although capable of modeling long-range correlations, have high computational complexity and are inefficient when processing high-resolution spatiotemporal data; Third, most existing methods, when fusing multiple meteorological elements, typically employ simple feature concatenation or weighted averaging, ignoring the complex nonlinear interactions and dynamic causal relationships between elements, resulting in insufficient information utilization; Fourth, in the face of extreme weather scenarios such as sudden cloud formations, heavy rain, and hail, traditional models lack mechanisms for identifying and adapting to the chaotic characteristics of the system, significantly increasing prediction errors; Fifth, purely data-driven models often fail to incorporate the physical constraints of photovoltaic power generation, potentially leading to prediction results that violate energy conservation, causality, and power boundary conditions, affecting the model's generalization ability and reliability. Therefore, there is an urgent need for an intelligent photovoltaic power prediction method that can integrate multi-scale spatiotemporal modeling, interactive analysis of meteorological elements, extreme event identification, and physical law constraints to improve prediction accuracy, efficiency, and robustness. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention proposes a multi-scale intelligent prediction method and system for photovoltaic power based on multiple spatiotemporal meteorological elements, thereby resolving the issues present in the prior art.

[0005] To achieve the above objectives, in a first aspect, the present invention provides a multi-scale intelligent prediction method for photovoltaic power based on multiple spatiotemporal meteorological elements, comprising: S1: Adaptive fractal spatiotemporal coding is performed on the input multivariate meteorological element time series data to generate a fractal embedding vector with scale awareness. S2: Input the fractal embedding vector into the wavelet neural operator to perform deep extraction and fusion of multi-scale frequency domain features in the wavelet domain; S3: Based on the multi-scale frequency domain features extracted by S2, a chaotic attention mechanism guided by the maximum Lyapunov exponent is adopted to adaptively enhance the modeling ability of key features under extreme weather conditions. S4: Input the optimized features from S3 into the physical information neural network, and perform physical consistency correction on the features by embedding and optimizing energy conservation, causality and power boundary constraints; S5: Construct a dynamic causal graph neural network from the features corrected in S4, model the time-varying interaction relationship of meteorological elements as nodes, and perform feature propagation and aggregation through spatiotemporal graph convolution; S6: Based on the characteristics of the fusion dynamic interaction relationship output by S5, a hierarchical Bayesian integration method is adopted to fuse the prediction results of multiple time scales, including ultra-short-term, short-term and medium-to-long-term, and to quantify the prediction uncertainty, outputting the final photovoltaic power prediction value and confidence interval.

[0006] Preferably, in step S1, the adaptive fractal spatiotemporal coding specifically involves: dynamically determining the number of decomposition layers of variational mode decomposition based on the Hausdorff dimension of meteorological time series, and generating fractal embedding vectors for each intrinsic mode function obtained from the decomposition.

[0007] Preferably, in step S1, the Hausdorff dimension is calculated using box counting, satisfying the formula: ; in, Let the dimension be Hausdorff; The formula for calculating the number of decomposition layers is as follows: ; The formula for calculating the fractal embedding vector is as follows: in, This is the IMF serial number.

[0008] Preferably, in step S2, the wavelet neural operator uses the Morlet wavelet basis function, as shown in the following formula: ; in, For scale parameters, These are translation parameters; The formula for convolution is: ; in, A learnable kernel function; The signal propagation formula for the multilayer wavelet neural operator is: ; in, For activation function, , For learnable parameters, This is a bias term.

[0009] Preferably, in step S3, the formula for calculating the maximum Lyapunov exponent is: ; in, The evolution of the initial perturbation; The weight calculation formula for the chaotic attention mechanism is as follows: ; in, For standard attention weights, To adjust the parameters, For the time span.

[0010] Preferably, in step S4, the formula for the energy conservation constraint is: ; in, For photoelectric conversion efficiency, For component area, for irradiance at any time For temperature coefficient, For component temperature, For reference temperature; The causal constraint is achieved through a causal mask. accomplish, when , when ; Physical boundary constraints are and , Rated power, This represents the maximum power change rate; The physical loss function is: .

[0011] Preferably, in step S5, the formula for the dynamic cause-effect graph is: ; in, For nodes; The nodes include irradiance, temperature, wind speed, humidity, cloud cover, and air pressure; The time-varying causal edge strength satisfies: ; The spatiotemporal graph convolution formula is: ; in For time-varying causal strength, For nodes The causal neighbor set, For the first layer node The hidden representation.

[0012] Preferably, in step S6, the different time scales include ultra-short-term scales, short-term scales, and medium-to-long-term scales; the prediction weights for each scale follow a Beta distribution and are dynamically updated based on historical prediction errors.

[0013] In a second aspect, the present invention also discloses a photovoltaic power multi-scale prediction system based on multiple spatiotemporal meteorological elements for implementing the method described in the first aspect, comprising: The data acquisition module is used to collect and preprocess multi-dimensional meteorological data. The fractal coding module, connected to the data acquisition module, is used to perform the adaptive fractal spatiotemporal coding; The wavelet operator module, connected to the fractal coding module, is used to implement the learning and mapping of wavelet neural operators; The chaotic attention module is connected to the data acquisition module and the wavelet operator module respectively, and is used to implement the chaotic attention mechanism. The physical constraint module is connected to the data acquisition module, the wavelet operator module, and the chaotic attention module, respectively, and is used to embed and calculate the physical constraints. The causal graph module, which is connected to the data acquisition module and the physical constraint module respectively, is used to construct a dynamic causal graph neural network and model the interaction of elements. An integrated prediction module, connected to the causal graph module, is used to perform hierarchical Bayesian integration and output prediction results and uncertainty intervals.

[0014] Thirdly, the present invention also discloses a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described in the first aspect.

[0015] Compared with the prior art, the present invention has the following advantages and technical effects: This invention achieves accurate, robust, and physically consistent prediction of photovoltaic power across all scales, from second-level to season-level, by constructing a sequential technical architecture of "fractal encoding—wavelet operator—chaotic attention—physical constraint—causal graph—Bayesian ensemble." Specifically, this scheme produces the following technical effects closely related to its technical features: 1. This invention employs a chaotic attention mechanism guided by the maximum Lyapunov exponent in step S3. This technical feature can dynamically identify the chaotic state of meteorological systems (corresponding to extreme weather) and adaptively modulate the attention weights, thereby enhancing the modeling of short-term abrupt changes. This enables the invention to effectively address scenarios where traditional models experience a sharp increase in error, such as sudden changes in cloud cover and torrential rain, and significantly reduces prediction errors under extreme weather conditions.

[0016] 2. This invention generates scale-aware feature representations through "adaptive fractal spatiotemporal coding" in step S1, combines it with the deep extraction and fusion of multi-scale frequency domain information through "wavelet neural operators" in step S2, and dynamically fuses multi-scale prediction results through "hierarchical Bayesian ensemble" in step S6. This series of technical features forms a complete multi-scale processing chain, enabling ultra-short-term, short-term, and medium-to-long-term predictions to share underlying features and collaboratively optimize, thus solving the problem of missing coupling effects caused by traditional methods that process different scale prediction tasks independently.

[0017] 3. This invention employs step S2, "deep extraction and fusion of multi-scale frequency domain features in the wavelet domain using wavelet neural operators," and utilizes fast wavelet transform to perform convolution operations. This technique enables the model to globally capture the long-distance dependency between meteorological elements and power in the frequency domain, while reducing computational complexity. This improves the model's training and inference efficiency while maintaining high-precision modeling capabilities, making it more suitable for engineering deployment and real-time prediction.

[0018] 4. This invention, through step S4, "constructing a physical information neural network and embedding and optimizing energy conservation, causality, and power boundary constraints," embeds the core physical prior knowledge of photovoltaic power generation into the data-driven model in the form of a differentiable loss function, forcing its output to satisfy energy conservation, temporal causality, and equipment operating boundaries. This effectively avoids the physically unreasonable predictions that may arise from a purely data-driven model, significantly improving the model's generalization ability and decision reliability in scenarios outside the training data distribution.

[0019] 5. This invention, through step S5, "constructing a dynamic causal graph to model the time-varying interaction relationships of meteorological elements," utilizes the Granger causality test as a foundation to dynamically learn the directional time-varying causal strength among elements such as irradiance, temperature, and wind speed, and employs graph convolution for feature propagation. This surpasses the simple splicing or static weighting of multiple meteorological elements in traditional methods, enabling a more comprehensive and rational utilization of the complex nonlinear interaction information between elements, providing deeper driving factors for power prediction.

[0020] 6. This invention, through step S6, "uses a hierarchical Bayesian ensemble method to fuse prediction results from multiple time scales, including ultra-short-term, short-term, and medium-to-long-term predictions, and quantifies prediction uncertainty." This technical feature not only intelligently fuses prediction results from multiple scales but also outputs a high-coverage prediction confidence interval through a Bayesian posterior distribution. This effect provides crucial risk assessment basis for power dispatching, energy storage control, and market transactions of photovoltaic power plants, achieving an upgrade in decision-making information from "single predicted value" to "predicted value and its credibility." Attached Figure Description

[0021] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart of the multi-scale intelligent prediction method for photovoltaic power based on multiple spatiotemporal meteorological elements according to an embodiment of the present invention. Figure 2 This is a diagram illustrating the overall architecture of the multi-scale intelligent prediction system according to an embodiment of the present invention. Figure 3 This is a flowchart of the Hausdorff dimension adaptive encoding process according to an embodiment of the present invention; Figure 4 This is a schematic diagram of the Lyapunov exponent-guided chaotic attention mechanism in an embodiment of the present invention. Figure 5 This is a diagram of the physical information neural network constraint framework of an embodiment of the present invention; Figure 6 This is a schematic diagram of hierarchical Bayesian integration and uncertainty quantification in an embodiment of the present invention; Figure 7 This is a schematic diagram of a multi-scale intelligent prediction system for photovoltaic power based on multiple spatiotemporal meteorological elements, according to an embodiment of the present invention. Detailed Implementation

[0022] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0023] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0024] Example 1 like Figure 1 As shown, this embodiment provides a multi-scale intelligent prediction method for photovoltaic power based on multiple spatiotemporal meteorological elements, including: S1: Adaptive fractal spatiotemporal coding based on the Hausdorff dimension of meteorological time series is performed to dynamically determine the number of decomposition layers of variational mode decomposition and generate fractal embedding vectors for each intrinsic mode function obtained by decomposition, providing a scale-adaptive spatiotemporal representation basis for subsequent multi-scale feature learning. Specifically, a fractal spatiotemporal representation is constructed based on Hausdorff dimension adaptive encoding, and the process is as follows: Figure 3 As shown, this includes: calculating the Hausdorff dimension of meteorological time series. To characterize the time complexity, based on Dynamically determine the number of decomposition levels for variational mode decomposition (VMD) For each intrinsic mode function (IMF) obtained from decomposition, a fractal embedding vector is generated to achieve scale-adaptive sensing. Specifically, this includes: 1.1 Calculate the temporal fractal dimension: For the input meteorological time series data Calculate the Hausdorff dimension The formula is: (1) in, The scale required to cover this time series data is The number of balls. The Hausdorff dimension can effectively characterize the time complexity: when When the weather is stable, the long-term trend dominates; when At that time, corresponding to scenarios such as cloud disturbances, multi-scale components are mixed; when At that time, corresponding to extreme weather, high-frequency components are significant.

[0025] 1.2 Adaptive Multiscale Decomposition: Based on Hausdorff dimension Dynamically select the number of decomposition layers in variational mode decomposition (VMD) The formula is: (2) in, ⌈・⌉ indicates rounding up. The meteorological time series signal is decomposed into... using VMD. Each intrinsic mode function (IMF) is a unique mode function.

[0026] 1.3 Fractal Embedding: Generate a fractal embedding vector for each IMF, using the following formula: (3) in, This is the IMF sequence number. This embedding vector enables the model to automatically perceive the scale characteristics of the current signal, providing a foundation for subsequent multi-scale predictions.

[0027] S2: Based on the fractal embedding vector generated by S1, a wavelet neural operator is used to learn the function space mapping from input meteorological elements to output power in the wavelet domain. Convolution operation is performed through fast wavelet transform to reduce computational complexity, and multiple layers are stacked to fuse multi-scale frequency domain information, thereby achieving deep extraction of multi-scale features in the fractal embedding vector. Specifically, the function space mapping is learned using wavelet neural operators (WNO), including: performing convolution operations in the wavelet domain and using Fast Wavelet Transform (FWT) to reduce the computational complexity to... Multi-layer WNO processing is used to process signals from different frequency bands and fuse multi-scale frequency domain information. Specifically, this includes: 2.1 Core Idea: Unlike the point-to-point mapping of traditional models, the Wavelet Neural Operator (WNO) learns the mapping operator G: G:A→U between function spaces, where A is the input function space and U is the output function space. The trained WNO can generalize to input data of arbitrary resolution without retraining.

[0028] 2.2 Wavelet Transform Convolution: Convolution is performed in the wavelet domain, with the following formula: (4) in, For Morlet wavelet basis functions, For scale parameters, For translation parameters, For learnable kernel functions, These are the weight parameters.

[0029] 2.3 Fast Wavelet Transform: The Fast Wavelet Transform (FWT) is used to reduce computational complexity, bringing the computational complexity of WNO down to a minimum. Much lower than the O(Transformer) n ²), while maintaining the global receptive field to effectively capture long-distance spatiotemporal dependencies.

[0030] 2.4 Multi-scale fusion: Stacking multiple WNO layers, each layer processes signals in different frequency bands. The signal propagation formula is: (5) in, For activation function, , For learnable parameters, This is the bias term. Through multi-level stacking, the final output contains complete timescale information in seconds, hours, and days.

[0031] S3: Based on the multi-scale frequency domain features output by S2, a chaotic attention mechanism guided by the maximum Lyapunov exponent is introduced. The system state is identified according to the maximum Lyapunov exponent, and the attention weight is adaptively modulated to enhance the attention to key frequency domain features in extreme weather scenarios. Specifically, such as Figure 4 The aforementioned Lyapunov exponent-guided chaotic attention (LCA) mechanism includes: calculating the maximum Lyapunov exponent of the meteorological time series. To identify the chaotic state of the system, based on Adaptive modulation of attention weights enhances short-term dependency modeling under extreme events. Specifically, this includes: 3.1 Lyapunov Index Calculation: Calculate the maximum Lyapunov index from meteorological time-series data. The formula is (6) in, This represents the evolution trajectory of the initial perturbation. The Lyapunov exponent can effectively identify the system state: when... At that time, the system is in a chaotic state, corresponding to extreme weather, and is highly sensitive to initial conditions; when At this time, the system is in a quasi-periodic state, corresponding to normal weather; when At that time, the system was in a stable state, corresponding to clear weather.

[0032] 3.2 Chaotic Attention Weights: Based on Adaptive modulation of attention weights, the formula is: (7) in, For standard attention weights, To adjust the parameters, The time span. This mechanism operates in a chaotic state ( Enhanced short-term dependency modeling under steady state ( ) By enhancing long-term dependency modeling, we can achieve adaptive identification and accurate prediction of extreme events, reducing extreme weather prediction errors by 45%.

[0033] S4: Based on the multi-scale frequency domain features optimized in S3, a physical information neural network constraint layer is constructed, embedding energy conservation constraints, causality constraints, and power physical boundary constraints. The model is optimized through a physical loss function to ensure that the features satisfy the physical laws of photovoltaic power generation. Specifically, such as Figure 5 As shown, a constraint layer for a physical information neural network is constructed, including embedding energy conservation constraints, causality constraints, and power physical boundary constraints. The model is optimized through a physical loss function to ensure that the output satisfies physical laws. Specifically, this includes: 4.1 Energy Conservation Constraint: Photovoltaic power output satisfies the energy conservation equation: (8) in, For photoelectric conversion efficiency, For component area, for irradiance at any time For temperature coefficient, For component temperature, This is a reference temperature.

[0034] Define the physical loss function: (9) The loss function constrains the model output to conform to the law of conservation of energy.

[0035] 4.2 Causality Constraint: Changes in photovoltaic power cannot precede changes in irradiance; therefore, a causal mask is introduced. : Ensuring that predictions rely solely on historical information, preventing the leakage of future information, and conforming to the law of causality.

[0036] 4.3 Boundary Condition Constraints: Photovoltaic power has physical boundaries. , .in, Rated power, This represents the maximum power change rate. By forcing the model output to satisfy the above boundary conditions through a projection layer, the model's generalization error is reduced by 30%.

[0037] S5: Based on the features of S4 after physical constraint correction, a dynamic causal graph neural network is established. Meteorological elements are used as nodes. The strength of time-varying causal edges is learned through Granger causality test. Based on spatiotemporal graph convolution propagation of node features, the dynamic interaction relationship between meteorological elements is modeled to further enrich the correlation information of features. Specifically, a Dynamic Causal Graph Neural Network (DCGNN) is established to model the interaction of meteorological elements. This includes: using meteorological elements as nodes, learning the strength of time-varying causal edges through Granger causality tests, propagating element features based on spatiotemporal graph convolution, and fusing information from multiple meteorological elements using attention pooling. This embodiment models the complex interactions between meteorological elements, making full use of diverse meteorological information. Specifically, it includes: 5.1 Cause-and-effect graph construction: Multiple meteorological elements, including irradiance, temperature, wind speed, humidity, cloud cover, and air pressure, are modeled as a dynamic cause-and-effect graph. The formula is: (10) in For a set of nodes, It is a time-varying set of causal edges.

[0038] The causal strength of an edge is learned through Granger causality testing and satisfies the formula: (11) That is, when removed After reading the historical information, If the prediction accuracy decreases significantly, then it is considered that... right Granger causality exists.

[0039] 5.2 Spatiotemporal Graph Convolution: Propagates information on a dynamic causal graph. The formula is: (12) in, The time-varying causal strength (learnable). For nodes The causal neighbor set, For the first layer node The hidden representation.

[0040] 5.3 Attention Pooling: Integrating features from multiple meteorological elements, the formula is as follows: (13) (14) in, Score the importance of nodes. This approach uses attention weights to automatically learn the importance of different meteorological elements, avoiding the subjectivity of manually setting weights.

[0041] S6: Based on the multi-scale features with dynamic interaction relationships output by S5, hierarchical Bayesian ensemble is adopted to configure dedicated prediction heads for different time scales. Multi-scale prediction results are fused based on dynamically updated weights, and prediction uncertainty is quantified through Bayesian posterior distribution to output the final photovoltaic power prediction value and confidence interval.

[0042] Specifically, such as Figure 6 As shown, this embodiment adopts hierarchical Bayesian ensemble fusion of multi-scale prediction results, and designs dedicated prediction heads for different time scales such as ultra-short-term, short-term, and medium-to-long-term. The prediction weights of each scale follow a Beta distribution and are dynamically updated according to historical errors. The prediction interval is estimated through Bayesian posterior distribution to achieve uncertainty quantification.

[0043] In this embodiment, multi-timescale prediction results are generated to achieve accurate prediction and uncertainty assessment. Specifically, this includes: 6.1 Multi-scale prediction head: Dedicated prediction heads are designed for different time scales: Ultra-short term (5-15 minutes): High-frequency IMF combined with chaotic attention is used to adapt to rapidly changing weather scenarios; Short term (1-4 hours): Mid-frequency IMF combined with wavelet operators is used to balance prediction accuracy and computational efficiency; Medium and long term (1-7 days): Low-frequency IMF combined with physical constraints is used to ensure prediction stability by utilizing long-term trends and physical laws.

[0044] 6.2 Bayesian Weight Learning: Prediction weights at each scale follow a Beta distribution. (15) in, , The prediction is dynamically updated based on historical prediction errors, making the weight allocation more closely match actual prediction needs. The final prediction result is: (16) in, .

[0045] 6.3 Uncertainty Quantification: Estimating the Prediction Interval via Bayesian Posterior Distribution: (17) This embodiment provides a 90% prediction range, with a range coverage of 92%, providing a risk assessment basis for photovoltaic power plant scheduling decisions.

[0046] Example 2 like Figure 7 As shown, this embodiment provides a photovoltaic power multi-scale prediction system based on multiple spatiotemporal meteorological elements for implementing the method described in Embodiment 1. The overall architecture of the system is as follows: Figure 2 As shown, the connection relationships and data flow paths of each functional module are clearly presented, including: The data acquisition module is used to collect and preprocess multi-dimensional meteorological data. Specifically, it collects real-time and historical data on diverse meteorological elements such as irradiance, temperature, wind speed, humidity, cloud cover, and air pressure.

[0047] The fractal coding module, connected to the data acquisition module, is used to perform the adaptive fractal spatiotemporal coding; Specifically, the Hausdorff dimension of the meteorological time series is calculated, the number of decomposition layers is dynamically determined, the intrinsic mode functions are obtained through variational mode decomposition, and fractal embedding vectors are generated.

[0048] The wavelet operator module, connected to the fractal coding module, is used to implement the learning and mapping of wavelet neural operators; Specifically, wavelet neural operators are used to learn function space mapping in the wavelet domain, fast wavelet transform is used to reduce computational complexity, and multi-layer stacking is used to achieve multi-band information fusion.

[0049] The chaotic attention module is connected to the data acquisition module and the wavelet operator module respectively, and is used to implement the chaotic attention mechanism. Specifically, the maximum Lyapunov exponent is calculated to identify chaotic states, and attention weights are modulated to optimize feature extraction.

[0050] The physical constraint module is connected to the data acquisition module, the wavelet operator module, and the chaotic attention module, respectively, and is used to embed and calculate the physical constraints. Specifically, energy conservation, causality, and boundary condition constraints are embedded to calculate physical losses and feed them back to model optimization.

[0051] The causal graph module, which is connected to the data acquisition module and the physical constraint module respectively, is used to construct a dynamic causal graph neural network and model the interaction of elements. Specifically, a dynamic causal graph is constructed, and meteorological element characteristics are propagated through spatiotemporal graph convolution to achieve multi-element interactive modeling.

[0052] An integrated prediction module, connected to the causal graph module, is used to perform hierarchical Bayesian integration and output prediction results and uncertainty intervals; Specifically, the integrated prediction module is equipped with dedicated prediction heads for different time scales, and multi-scale prediction results are fused through hierarchical Bayesian integration to output the final prediction value and 90% prediction interval.

[0053] Example 3 This embodiment uses a 100MW mountain photovoltaic power station multi-scale prediction as an example to verify the scheme described in Embodiment 2. Specifically, it includes: Scenario Description: The photovoltaic power station is located in a mountainous area with complex terrain, violent cloud movement, and variable weather conditions. It needs to achieve photovoltaic power prediction on a full time scale of 15 minutes to 7 days to support the optimized scheduling of the power station.

[0054] System Configuration: Input meteorological elements: irradiance, temperature, wind speed, humidity, cloud cover, and air pressure; data sampling frequency is 5 minutes. Fractal coding module: Hausdorff dimension is calculated using box counting, and the number of VMD decomposition layers is dynamically adjusted according to D_H (range 5-9 layers). Wavelet operator module: 6-layer WNO, each layer processes 1 octave, using Morlet wavelet basis functions, and achieving O(nlogn) complexity through fast wavelet transform; Chaotic attention module: Adjust parameter γ=0.3 to calculate Lyapunov index in real time to identify extreme weather; Physical constraint module: photoelectric conversion efficiency η = 0.22, component area A = 5.5 × 10⁻⁶ 4 m², temperature coefficient β=0.004 / ℃, reference temperature T_ref=25℃, rated power P_rated=100MW, maximum power change rate R_max=10MW / min; Causal graph module: 6 meteorological element nodes, learn time-varying causal strength through Granger causality test, spatiotemporal graph convolution layer has 3 layers; Integrated prediction module: Dedicated prediction heads for ultra-short-term (15 minutes), short-term (4 hours), and medium-to-long-term (1 day, 7 days). The initial values ​​of the weighted Beta distribution parameters α_k and β_k are both 1, and they are updated every 24 hours based on historical errors.

[0055] Training data: Two years of historical meteorological data and photovoltaic power data from the power station were used, with 80% used for training, 10% for verification, and 10% for testing.

[0056] Example 4 This embodiment uses extreme weather prediction for a 50MW plain photovoltaic power station as an example to verify the scheme described in Embodiment 2. Specifically, it includes: Scene characteristics: The photovoltaic power station is located in a plain area, where there are frequent thunderstorms, hail and other extreme weather in summer, which requires high robustness of the forecasting system and the ability to predict extreme events.

[0057] Implementation Highlights: Real-time monitoring of the Lyapunov index: When λ_max > 0.5 is detected, it automatically switches to extreme weather prediction mode to enhance the short-term dependency modeling of chaotic attention; Physical constraint protection: Under extreme weather conditions, strengthen power boundary constraints and energy conservation constraints to prevent the model output from exceeding the physically reasonable range of predicted values; Dynamic adjustment of cause-effect diagram: During heavy rain, the weight of the cause-effect chain "air pressure → cloud cover → irradiance" is automatically increased to enhance the influence weight of key elements; Uncertainty quantification: Under extreme weather conditions, the forecast interval is automatically expanded to ensure that the coverage of the forecast interval remains above 90%.

[0058] Example 5 This embodiment also discloses a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the method described in Embodiment 1.

[0059] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A multi-scale intelligent prediction method for photovoltaic power based on multiple spatiotemporal meteorological elements, characterized in that, Includes the following steps: S1: Adaptive fractal spatiotemporal coding is performed on the input multivariate meteorological element time series data to generate a fractal embedding vector with scale awareness. S2: Input the fractal embedding vector into the wavelet neural operator to perform deep extraction and fusion of multi-scale frequency domain features in the wavelet domain; S3: Based on the multi-scale frequency domain features extracted by S2, a chaotic attention mechanism guided by the maximum Lyapunov exponent is adopted to adaptively enhance the modeling ability of key features under extreme weather conditions. S4: Input the optimized features from S3 into the physical information neural network, and perform physical consistency correction on the features by embedding and optimizing energy conservation, causality and power boundary constraints; S5: Construct a dynamic causal graph neural network from the features corrected in S4, model the time-varying interaction relationship of meteorological elements as nodes, and perform feature propagation and aggregation through spatiotemporal graph convolution; S6: Based on the characteristics of the fusion dynamic interaction relationship output by S5, a hierarchical Bayesian integration method is adopted to fuse the prediction results of multiple time scales, including ultra-short-term, short-term and medium-to-long-term, and to quantify the prediction uncertainty, outputting the final photovoltaic power prediction value and confidence interval.

2. The method according to claim 1, characterized in that, In step S1, the adaptive fractal spatiotemporal coding specifically involves: dynamically determining the number of decomposition layers of variational mode decomposition based on the Hausdorff dimension of meteorological time series, and generating fractal embedding vectors for each intrinsic mode function obtained from the decomposition.

3. The method according to claim 2, characterized in that, In step S1, the Hausdorff dimension is calculated using box counting, satisfying the formula: ； in, Let the dimension be Hausdorff; The formula for calculating the number of decomposition layers is as follows: ； The formula for calculating the fractal embedding vector is as follows: ； in, This is the IMF serial number.

4. The method according to claim 1, characterized in that, In step S2, the wavelet neural operator uses the Morlet wavelet basis function, as shown in the formula: ； in, For scale parameters, These are translation parameters; The formula for convolution is: ； in, For learnable kernel functions; The signal propagation formula for the multilayer wavelet neural operator is: ； in, For activation function, , For learnable parameters, This is a bias term.

5. The method according to claim 1, characterized in that, In step S3, the formula for calculating the maximum Lyapunov exponent is as follows: ； in, The evolution of the initial perturbation; The weight calculation formula for the chaotic attention mechanism is as follows: ； in, For standard attention weights, To adjust the parameters, For the time span.

6. The method according to claim 1, characterized in that, In step S4, the formula for the energy conservation constraint is: ； in, For photoelectric conversion efficiency, For component area, for irradiance at any time For temperature coefficient, For component temperature, For reference temperature; The causal constraint is achieved through a causal mask. accomplish, when , when ; Physical boundary constraints are and , Rated power, This represents the maximum power change rate. The physical loss function is: 。 7. The method according to claim 1, characterized in that, In step S5, the time-varying interaction relationship is learned through Granger causality test to learn the causal edge strength. The formula for the dynamic causal graph neural network is: ； in, For nodes; The nodes include irradiance, temperature, wind speed, humidity, cloud cover, and air pressure; The time-varying causal edge strength satisfies: ； The spatiotemporal graph convolution formula is: ； in For time-varying causal strength, For nodes The causal neighbor set, For the first layer node The hidden representation.

8. The method according to claim 1, characterized in that, In step S6, the different time scales include ultra-short-term scale, short-term scale, and medium-to-long-term scale; the prediction weights for each scale follow a Beta distribution and are dynamically updated based on historical prediction errors.

9. A photovoltaic power multi-scale prediction system based on multiple spatiotemporal meteorological elements for implementing the method of any one of claims 1 to 8, characterized in that, include: The data acquisition module is used to collect and preprocess multi-dimensional meteorological data. The fractal coding module, connected to the data acquisition module, is used to perform the adaptive fractal spatiotemporal coding; The wavelet operator module, connected to the fractal coding module, is used to implement the learning and mapping of wavelet neural operators; The chaotic attention module is connected to the data acquisition module and the wavelet operator module respectively, and is used to implement the chaotic attention mechanism. The physical constraint module is connected to the data acquisition module, the wavelet operator module, and the chaotic attention module, respectively, and is used to embed and calculate the physical constraints. The causal graph module, which is connected to the data acquisition module and the physical constraint module respectively, is used to construct a dynamic causal graph neural network and model the interaction of elements. An integrated prediction module, connected to the causal graph module, is used to perform hierarchical Bayesian integration and output prediction results and uncertainty intervals.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the steps of the method according to any one of claims 1-8.

Images