Photovoltaic power prediction method fusing spatial-aware regression and cross-view knowledge transfer

Abstract

A photovoltaic power prediction method integrating spatial perception regression and cross-perspective knowledge transfer includes: S1. Acquiring image sequence data and preprocessing it; S2. Constructing a feature extraction architecture and a spatial semantic mapping and regression network, with the spatial semantic mapping and regression network connected to the output of the target spatiotemporal feature extraction network; S3. Using a reference high-resolution feature extraction network as a supervision source, projecting the intermediate layer features of the target spatiotemporal feature extraction network onto the reference feature space, calculating the multi-scale composite distillation loss, and constructing a joint objective function by combining it with the power prediction loss, and coordinating the parameter updates of the target spatiotemporal feature extraction network and the spatial semantic mapping and regression network; S4. Blocking the data path of the reference high-resolution feature extraction network, inputting the first-resolution image sequence of the test set into the trained target spatiotemporal feature extraction network, and transmitting it to the spatial semantic mapping and regression network to directly output the photovoltaic power prediction value.

Description

Technical Field

[0001] This invention relates to the fields of renewable energy management and computer vision, and in particular to a method for short-term photovoltaic power prediction using deep learning technology through cross-view (satellite-ground) feature alignment and spatial semantic mapping. Background Technology

[0002] Solar energy is central to global decarbonization efforts, but its inherent intermittency poses a significant challenge to grid stability. Cloud-induced power generation fluctuations can cause frequency instability and increase operating costs; therefore, accurate short-term photovoltaic power forecasting is crucial.

[0003] Existing photovoltaic (PV) forecasting technologies suffer from a typical trade-off between coverage and resolution at the data source level: Satellite data (such as geostationary meteorological satellites like Himawari-8): While offering broad coverage and low acquisition costs, its spatial resolution is relatively coarse (typically 1-4 kilometers), making it difficult to distinguish the dynamic characteristics of tiny cloud clusters that drastically affect the PV irradiance of individual stations. Ground-based all-sky imagers (ASI): Although providing meter-level high-resolution cloud images and accurately capturing local obstructions, their field of view is extremely limited, and the hardware deployment and maintenance costs are high, hindering large-scale deployment.

[0004] At the model architecture level, existing technologies have not yet proposed an effective mechanism to solve the problem of heterogeneous data fusion. Existing end-to-end regression methods often treat the network as a "black box," lacking explicit modeling of the physical meaning of cloud spatial distribution.

[0005] The technical challenge lies in how to enable a scalable model based on low-resolution satellite data to perceive and utilize the fine-grained cloud dynamics that only high-resolution ground data can provide, without actually using expensive and limited-coverage ground data during inference. In other words, how to creatively combine existing spatiotemporal prediction architectures and feature extraction architectures to create a synergistic effect, thereby resolving the aforementioned contradiction between resolution and coverage. Summary of the Invention

[0006] This invention aims to address the shortcomings of existing technologies, such as coarse satellite data resolution, small ground data coverage, and unstable multimodal data fusion, by providing a photovoltaic power prediction method that integrates spatial perception regression and cross-perspective knowledge transfer.

[0007] This invention constructs an asymmetric parallel feature extraction architecture, utilizes a multi-level cross-domain feature alignment mechanism to transfer high-precision prior knowledge from ground observations to satellite data streams, and combines a spatial perception regression module to generate physically meaningful intermediate representations, thereby achieving low-cost, high-precision, and physically interpretable photovoltaic power prediction.

[0008] The first aspect of this invention provides a photovoltaic power prediction method integrating spatially perceived regression and cross-perspective knowledge transfer. The method includes the following steps: S1. Acquiring image sequence data, including a low-frequency wide-field-of-view first-resolution image sequence, a high-frequency narrow-field-of-view second-resolution image sequence, and historical power data of a photovoltaic power plant; preprocessing all data sequences to establish an aligned input stream based on geographic coordinate anchor points, and dividing the data into training, validation, and test sets; S2. Constructing a feature extraction architecture and a spatial semantic mapping and regression network; the feature extraction architecture includes a target spatiotemporal feature extraction network and a reference high-resolution feature extraction network; the spatial semantic mapping and regression network is connected to the output of the target spatiotemporal feature extraction network; S3. The training set data is input into the feature extraction architecture. The spatiotemporal feature extraction network outputs spatiotemporal features to the spatial semantic mapping and regression network. After decoding, a two-dimensional spatial influence weight map is generated, and a preliminary photovoltaic power prediction value is output through a global weighted aggregation operation. The power prediction loss is calculated based on the preliminary photovoltaic power prediction value and the historical power data. At the same time, using the reference high-resolution feature extraction network as the supervision source, the intermediate layer features of the target spatiotemporal feature extraction network are projected to the reference feature space through the feature dimension adaptation and mapping unit. The multi-scale composite distillation loss, which includes semantic consistency and spatial saliency distribution differences, is calculated. A joint objective function including the power prediction loss and the multi-scale composite distillation loss is constructed to collaboratively update the parameters of the target spatiotemporal feature extraction network and the spatial semantic mapping and regression network. S4. In the inference stage, the data path of the reference high-resolution feature extraction network is blocked. The first resolution image sequence of the test set is input into the trained target spatiotemporal feature extraction network to infer the spatiotemporal features containing high-frequency texture information, and transmits them to the spatial semantic mapping and regression network to directly output the final photovoltaic power prediction value.

[0009] Further, step S1 includes:

[0010] S11: Data source configuration: Collect geostationary meteorological satellite cloud image data as the first resolution image sequence; collect ground all-sky imager data as the second resolution image sequence; collect the actual power generation of the power station as the regression target.

[0011] S12: Preprocessing and Alignment: Perform center alignment based on geographic coordinates on the first resolution image sequence and the second resolution image sequence, and perform independent statistical normalization on the image channel data and power data.

[0012] S13 Data Cleaning and Dataset Division: Remove invalid samples from nighttime and those missing from sensors, and divide the processed sequence data into training set, validation set and test set.

[0013] Further, the target spatiotemporal feature extraction network in step S2 is configured to process a first-resolution image sequence with low frequency and wide field of view; the reference high-resolution feature extraction network is configured to process a second-resolution image sequence with high frequency and narrow field of view; and a center alignment process based on geographic coordinates is performed on the first-resolution image sequence and the second-resolution image sequence.

[0014] Furthermore, the target spatiotemporal feature extraction network and the reference high-resolution feature extraction network mentioned in step S2 both adopt a hierarchical spatiotemporal prediction architecture; the architecture connects the spatial encoder, spatiotemporal converter and spatial decoder sequentially from the input end to the output end.

[0015] The spatial encoder consists of multiple stacked convolutional layers, used to extract spatial features and reduce resolution; the spatiotemporal converter consists of multiple stacked Inception modules based on multi-scale parallel convolution, used to capture spatiotemporal evolution patterns while maintaining the spatial resolution of the feature map; and the spatial decoder consists of multiple sequentially stacked deconvolutional layers, used to restore spatial dimensions.

[0016] The target spatiotemporal feature extraction network is configured as a lightweight architecture with fewer spatiotemporal converter layers to process image sequences at the first resolution; the reference high-resolution feature extraction network is configured as a deep architecture with more spatiotemporal converter layers to process image sequences at the second resolution.

[0017] The spatial semantic mapping and regression network adopts an encoder-decoder structure with skip connections. The decoder upsamples the features and fuses the shallow spatial features from the encoder through the skip connections. At the end of the decoder, the feature map is compressed into a single-channel two-dimensional spatial influence weight map. The pixel values ​​of this map represent the local influence weight of cloud cover on photovoltaic power at the corresponding geographical location.

[0018] By using a learnable weight tensor to perform weighted summation and bias correction on the two-dimensional spatial influence weight map, the scalar photovoltaic power prediction value is calculated.

[0019] Furthermore, the spatial semantic mapping and regression network specifically includes: an input adaptation layer: adjusting the number of channels of the input features using convolutional layers; a shrinking path: composed of multiple convolutional blocks and downsampling layers cascaded together, used to extract deep semantic features and reduce resolution; an expanding path: composed of multiple upsampling layers and convolutional blocks cascaded together, where the feature map after each upsampling layer is concatenated with the feature map from the corresponding layer of the shrinking path to restore spatial resolution; a regression generation layer: located at the end of the expanding path, configured to compress the multi-channel feature map into a single-channel two-dimensional spatial influence weight map through convolutional layers, which represents the local influence weight of the sky conditions at the corresponding geographical location on photovoltaic power; and a global weighted aggregation layer, configured to flatten the two-dimensional spatial influence weight map and use learnable weight parameters to perform weighted summation and bias correction on the flattened vector, directly mapping the output scalar photovoltaic power prediction value.

[0020] Further, step S3 includes:

[0021] S31: Adapting Feature Dimensions: For multiple corresponding depth levels of the target spatiotemporal feature extraction network and the reference high-resolution feature extraction network, an adaptation module composed of a bilinear upsampling layer and a convolutional layer is used to project the intermediate feature tensor of the target spatiotemporal feature extraction network in terms of spatial size and channel dimension to be consistent with the reference high-resolution feature extraction network.

[0022] S32: Calculate the multi-scale composite distillation loss and preliminary power prediction loss: Calculate the L2 norm squared distance between the adapted target feature tensor and the reference feature tensor as the semantic consistency loss; aggregate the activation value amplitudes of the feature tensors along the channel dimension to generate a spatial saliency map, and calculate the L1 norm distance between the two spatial saliency maps as the spatial saliency alignment loss; weight and sum the semantic consistency loss and spatial saliency distribution difference loss at each level to obtain the multi-scale composite distillation loss; input the training set data into the feature extraction architecture, use the target spatiotemporal feature extraction network to output spatiotemporal features to the spatial semantic mapping and regression network, decode to generate a two-dimensional spatial influence weight map, and output the preliminary photovoltaic power prediction value through a global weighted aggregation operation; calculate the power prediction loss based on the preliminary photovoltaic power prediction value and the historical power data;

[0023] S33: Construct a joint objective function, weighted by the photovoltaic power prediction loss, cloud prediction reconstruction loss, and the multi-scale composite distillation loss; the power prediction loss uses mean squared error and is configured as a physical quantity-guided regularization term; during the collaborative parameter update process, the gradient signal generated by the power prediction loss is used to back-calibrate the response weights of the target spatiotemporal feature extraction network to the input features. The cloud prediction reconstruction loss uses an L1 loss function to constrain the pixel-level difference between the predicted cloud map and the actual cloud map. The joint objective function is defined as:

[0024] (4)

[0025] in Predict reconstruction losses for cloud computing. For photovoltaic power prediction loss, This represents the loss from multi-scale complex distillation. To balance the hyperparameters.

[0026] Furthermore, the multi-scale composite distillation loss calculation process described in step S32 is as follows:

[0027] 1) Semantic consistency (feature mimicry): This involves calculating the adapted target feature map. Compared with reference feature map This is achieved by using the squared L2 norm distance between them:

[0028] (1)

[0029] in, Indicates the first Semantic consistency loss at each supervisory level; This indicates that the adapted target spatiotemporal feature extraction network is in the [missing information - likely a specific phase or stage]. Intermediate feature tensors of the layer; This indicates that the reference high-resolution feature extraction network is in the corresponding... The intermediate feature tensor of the layer (this feature comes from the frozen teacher network). This represents the square of the L2 norm.

[0030] 2) Spatial saliency alignment (attention mimicry): First, the amplitude of feature activation is aggregated along the channel axis to generate a spatial attention map. Then calculate the L1 norm distance between the two attention maps:

[0031] (2)

[0032] in, Indicates the first Spatial saliency alignment loss at each supervisory level; Represents the spatial attention graph of the target network; This represents the spatial attention graph of the reference network. This represents the L1 norm.

[0033] 3) Total losses from multi-scale combined distillation: For The losses of each supervisory level are weighted and summed:

[0034] (3)

[0035] , This represents the calculated total multiscale composite distillation loss; This indicates the total number of network levels involved in distillation supervision; Indicates the first Weighted balancing coefficients for hierarchical semantic consistency loss; Indicates the first Weighted balance coefficients for hierarchical spatial saliency alignment loss.

[0036] Further, step S4 includes:

[0037] S41: Single-stream deployment configuration: During the model deployment phase, the reference high-resolution feature extraction network is no longer loaded;

[0038] S42: End-to-end feature mapping and prediction: The first resolution image sequence of the test set is input into the target spatiotemporal feature extraction network, a forward convolution operation is performed to output a spatiotemporal feature tensor, and the spatiotemporal feature tensor is input into the spatial semantic mapping and regression network; the spatial semantic mapping and regression network generates the two-dimensional spatial influence weight map based on the received spatiotemporal feature tensor, and uses the learnable weight tensor to perform global weighted summation and bias correction on the map to obtain the predicted power value.

[0039] Further, the global weighted aggregation in step S42 includes: [the process of] the weight graph... Flattening is achieved through a fully connected layer, i.e., weighted summation. The regression analysis yields the final predicted photovoltaic power value:

[0040] (5)

[0041] in, Indicates at time The predicted scalar photovoltaic power value; Represents the pixel coordinate index of the two-dimensional spatial feature map; Indicates at time The generated two-dimensional spatial influence weight map in coordinates The pixel value at the location represents the weight of the local impact of cloud cover on photovoltaic power at the corresponding geographical location. This represents a learnable global weight parameter with the same size as the weighted graph, used to weight the influence of different spatial locations; This represents the bias correction term for the regression layer.

[0042] The second aspect of the present invention relates to a photovoltaic power prediction system that integrates spatial perception regression and cross-perspective knowledge transfer, comprising: an asymmetric parallel spatiotemporal feature extraction architecture construction module, used to construct a target spatiotemporal feature extraction network and a reference high-resolution feature extraction network respectively, and to establish a multi-source image sequence input data stream based on geographic coordinate anchor points;

[0043] The feature dimension adaptation and multi-level cross-domain alignment execution module is used to use the reference high-resolution feature extraction network as the supervision source, and project the intermediate layer features of the target spatiotemporal feature extraction network to the reference feature space through the feature dimension adaptation and mapping unit, and calculate the multi-scale composite distillation loss including semantic consistency and spatial saliency distribution differences.

[0044] The spatial influence weight intermediate representation and power mapping generation module is used to input the spatiotemporal features output by the target spatiotemporal feature extraction network into the spatial semantic mapping and regression network, generate a two-dimensional spatial influence weight map after decoding, and output the photovoltaic power prediction value through global weighted aggregation operation.

[0045] The power numerical feedback-driven end-to-end collaborative calibration execution module is used to construct a joint objective function that includes spatiotemporal reconstruction error, multi-scale composite distillation error and power regression error, and uses the power numerical error as a regularization signal for collaborative parameter updates.

[0046] The single-path independent inference execution module blocks the data path of the reference high-resolution feature extraction network during the inference phase and performs photovoltaic power prediction only based on the target spatiotemporal feature extraction network.

[0047] The innovation of this invention lies in:

[0048] 1. An asymmetric cross-view feature transfer architecture was constructed: breaking through the dependence of traditional multimodal fusion on dual-end input, an asymmetric structure of “deep reference network (ground)” guiding “shallow target network (satellite)” was designed, and the mismatch problem of heterogeneous data in spatial resolution and channel dimension was solved by feature dimension adaptation unit.

[0049] 2. A multi-scale distillation mechanism based on spatial saliency alignment is proposed: dual constraints are introduced into the shallow, middle and deep layers of the network. By simultaneously minimizing the semantic manifold distance and the difference in spatial saliency distribution, the low-resolution network is forced to implicitly reconstruct fine cloud edge textures in the absence of high-frequency inputs.

[0050] 3. A spatially perceptive power regression network with physical interpretability was designed: abandoning the traditional black-box fully connected regression, the decoder was innovatively used to generate a "two-dimensional spatial influence weight map" as an intermediate physical representation.

[0051] 4. A closed-loop calibration strategy for visual features guided by physical quantities is introduced: using the numerical error of photovoltaic power as a physical regularization term, and through the backpropagation dynamic calibration network for the screening mechanism of visual features, the optimization logic of using physical results to suppress background visual noise is realized.

[0052] The advantages of this invention are:

[0053] 1. Balancing low cost and high precision: This invention completely eliminates the reliance on expensive ground-based all-sky imagers during the reasoning stage, requiring only low-cost satellite data to operate. At the same time, through knowledge transfer, it gains the ability to capture tiny cloud clusters, similar to ground-based observation equipment, effectively resolving the trade-off between "coverage" and "resolution".

[0054] 2. Significantly improves the physical interpretability of the model: The generated two-dimensional spatial influence weight map intuitively shows the specific impact area of ​​cloud cover on power, providing grid dispatchers with a visualized decision-making basis and overcoming the lack of transparency in traditional end-to-end deep learning models.

[0055] 3. Enhance the robustness of heterogeneous data fusion: By adopting a soft alignment strategy in the feature space, the geometric registration problem caused by the difference between satellite top view and ground top view is effectively avoided, and it has a stronger generalization ability compared with direct stitching fusion.

[0056] 4. Achieve efficient edge deployment: Based on the single-path independent inference mechanism, the system does not need to load a large reference network during actual operation, significantly reducing computational overhead and making it extremely suitable for deployment on edge computing terminals in photovoltaic power plants with limited computing power. Attached Figure Description

[0057] Figure 1 The flowchart below shows the overall architecture of the photovoltaic power prediction method in Embodiment 1 of the present invention, illustrating the logical relationship between asymmetric parallel training and single-path independent inference.

[0058] Figure 2This is a schematic diagram of the internal topology of the target spatiotemporal feature extraction network (based on SimVP architecture) in Embodiment 1 of the present invention, showing the topology of the spatial encoder, spatiotemporal converter and spatial decoder.

[0059] Figures 3(a)-3(c) are comparison diagrams of cloud field prediction effects at different time steps in the embodiments of the present invention, demonstrating the effect of implicit feature reconstruction; wherein Figure 3(a) is the input cloud map, and Figures 3(b) and 3(c) are the predicted cloud maps at different time steps (step 8 and step 14, respectively).

[0060] Figures 4(a) and 4(b) are visualization heatmaps of the two-dimensional spatial influence weight map in the embodiment of the present invention, showing the local weight distribution of the influence of cloud cover on power. Figure 4(a) is a cloud field image at a certain moment, and Figure 4(b) is the corresponding spatial regression weight map represented by black and white shaded lines.

[0061] Figure 5 The diagram shows the functional modules of a photovoltaic power prediction system that integrates spatial perception regression and cross-perspective knowledge transfer according to the present invention. It illustrates the logical connections and data flow of five core modules: asymmetric architecture construction, cross-domain alignment, power mapping generation, collaborative calibration, and single-path inference. Detailed Implementation

[0062] To make the objectives, technical solutions, and advantages of the present invention clearer, embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Those skilled in the art should understand that the embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0063] Example 1

[0064] This embodiment provides a photovoltaic power prediction method that integrates spatial perception regression and cross-perspective knowledge transfer.

[0065] The method disclosed in this embodiment is preferably implemented under the PyTorch framework and executed on a high-performance computing device (such as an NVIDIA RTX 4090 GPU). To verify the effectiveness of the invention, this embodiment constructs an experimental environment based on real power plant data. The specific implementation process is carried out according to the aforementioned steps S1 to S4:

[0066] Step S1 collects image sequence data, including a low-frequency wide field of view first resolution image sequence, a high-frequency narrow field of view second resolution image sequence, and historical power data of photovoltaic power plants; preprocesses all data sequences to establish an aligned input stream based on geographic coordinate anchor points, and divides them into training set, validation set, and test set.

[0067] In this embodiment, the specific details are as follows:

[0068] S11 data source configuration: Collect geostationary meteorological satellite cloud image data as the first resolution image sequence; collect ground all-sky imager data as the second resolution image sequence; collect the actual power generation of the power plant as the regression target.

[0069] First-resolution image sequence (target input): Observational data from geostationary meteorological satellites (e.g., Himawari-8 satellite) are selected. Six bands covering visible and infrared light are chosen, specifically including the 0.65 μm, 0.86 μm, 3.9 μm, 7.0 μm, 11.2 μm, and 12.3 μm bands, with a time sampling interval set to 10 minutes.

[0070] Second resolution image sequence (reference input): Select high-resolution RGB images acquired by the All-Sky Imager (ASI) deployed at the photovoltaic power plant site, with a time interval of 10 minutes.

[0071] Historical photovoltaic power data (monitoring signal): Inverter output power data of the target photovoltaic power plant (e.g., the solar research facility in Gatton, Australia, with a total installed capacity of 3.275 MWp) are collected. The raw data resolution may be high (e.g., minute-level), and needs to be processed into a sequence corresponding to image frames.

[0072] S12 Preprocessing and Alignment: Perform center alignment based on geographic coordinates on the first resolution image sequence and the second resolution image sequence, and perform independent statistical normalization on the image channel data and power data.

[0073] Spatial alignment: Using the geographical coordinates of the photovoltaic power station as anchor points, image patches covering the area surrounding the power station are cropped from the overall satellite data. In this embodiment, all six bands are resampled to a unified standard using bilinear interpolation. Pixel resolution ( ).

[0074] Normalization: Independent channel normalization (Z-Score) is performed on all channels of the satellite image using the mean and standard deviation calculated from the training set. Power values ​​are normalized (e.g., max-min normalization) to eliminate the influence of dimensions and facilitate network regression convergence.

[0075] S13 Data Cleaning and Dataset Partitioning:

[0076] After removing invalid samples from nighttime and those missing from the sensors, the processed sequence data is divided according to a preset ratio (e.g., training set: validation set: test set) for subsequent model training and evaluation.

[0077] Sequence Construction: In this example, a 10-minute time interval is selected to construct a historical input sequence of 14 frames (covering 140 minutes). ) and 14 frames of future predicted target sequence ( ).

[0078] Step S2: Construct a feature extraction architecture and a spatial semantic mapping and regression network; the feature extraction architecture includes a target spatiotemporal feature extraction network and a reference high-resolution feature extraction network; the spatial semantic mapping and regression network is connected to the output of the target spatiotemporal feature extraction network;

[0079] Both the target spatiotemporal feature extraction network and the reference high-resolution feature extraction network adopt a hierarchical spatiotemporal prediction architecture; the architecture connects the spatial encoder, spatiotemporal converter and spatial decoder sequentially from the input end to the output end.

[0080] The spatial encoder consists of multiple stacked convolutional layers, used to extract spatial features and reduce resolution; the spatiotemporal converter consists of multiple stacked Inception modules based on multi-scale parallel convolution (the core component of GoogLeNet, which extracts features in parallel through multiple convolutional kernels of different sizes and performs channel concatenation), used to capture spatiotemporal evolution patterns while maintaining the spatial resolution of the feature map; the spatial decoder consists of multiple sequentially stacked deconvolutional layers, used to restore spatial dimensions.

[0081] The target spatiotemporal feature extraction network is configured as a lightweight architecture with fewer spatiotemporal converter layers to process image sequences at the first resolution; the reference high-resolution feature extraction network is configured as a deep architecture with more spatiotemporal converter layers to process image sequences at the second resolution.

[0082] Specifically, in this embodiment, the specific parameter design of the target spatiotemporal feature extraction network (student network) is shown in Table 1. Table 1 is the parameter configuration table for the target spatiotemporal feature extraction network (student network). Taking full account of the low-resolution characteristics (40x40) of satellite cloud images, the multi-scale convolutional kernels of the Inception module (such as...) are utilized. , , This approach simultaneously captures the generation of small cloud clusters and their translational motion through large-scale cloud clusters. It is similar in structure to the high-resolution feature extraction network (teacher network), but its input size is... Furthermore, the number of Inception modules stacked in the middle is increased to 16-24 layers to extract deeper texture features.

[0083] Table 1

[0084]

[0085] The spatial semantic mapping and regression network adopts an encoder-decoder structure with skip connections. The decoder upsamples the features and fuses the shallow spatial features from the encoder through the skip connections. At the end of the decoder, the feature map is compressed into a single-channel two-dimensional spatial influence weight map. The pixel values ​​of this map represent the local influence weight of cloud cover on photovoltaic power at the corresponding geographical location.

[0086] By using a learnable weight tensor to perform weighted summation and bias correction on the two-dimensional spatial influence weight map, the scalar photovoltaic power prediction value is calculated.

[0087] Specifically, the system includes: an input adaptation layer that adjusts the number of channels in the input features using convolutional layers; a shrinking path composed of cascaded convolutional blocks and downsampling layers, used to extract deep semantic features and reduce resolution; an expanding path composed of cascaded upsampling layers and convolutional blocks, where the feature map after each upsampling layer is concatenated with the feature map from the corresponding layer of the shrinking path to restore spatial resolution; a regression generation layer located at the end of the expanding path, configured to compress the multi-channel feature map into a single-channel two-dimensional spatial influence weight map through convolutional layers, which represents the local influence weight of the sky conditions at the corresponding geographical location on photovoltaic power; and a global weighted aggregation layer configured to flatten the two-dimensional spatial influence weight map and use learnable weight parameters to perform weighted summation and bias correction on the flattened vector, directly mapping the output scalar photovoltaic power prediction value.

[0088] In this embodiment, the network (Image2PV Module) is designed to establish a nonlinear mapping between cloud spatial distribution and power values. Specific hierarchical connections and parameters are shown in Table 2, which is the parameter configuration table for the Spatial Semantic Mapping and Regression Network (Image2PV).

[0089] Table 2

[0090]

[0091] Furthermore, the construction process also includes: designing the bias term after the regression generation layer based on the installed capacity of the photovoltaic power station and the distribution range of historical power data. The initial values ​​are set to accelerate the convergence speed of the model in the early stages of training.

[0092] Step S3: Input the training set data into the feature extraction architecture, use the target spatiotemporal feature extraction network to output spatiotemporal features to the spatial semantic mapping and regression network, decode to generate a two-dimensional spatial influence weight map, and output the preliminary photovoltaic power prediction value through global weighted aggregation operation; calculate the power prediction loss based on the preliminary photovoltaic power prediction value and the historical power data; at the same time, using the reference high-resolution feature extraction network as the supervision source, project the intermediate layer features of the target spatiotemporal feature extraction network to the reference feature space through the feature dimension adaptation and mapping unit, calculate the multi-scale composite distillation loss including semantic consistency and spatial saliency distribution differences, construct a joint objective function including the power prediction loss and the multi-scale composite distillation loss, and perform collaborative parameter updates for the target spatiotemporal feature extraction network and the spatial semantic mapping and regression network;

[0093] S31: Adapting Feature Dimensions: For multiple corresponding depth levels of the target spatiotemporal feature extraction network and the reference high-resolution feature extraction network, an adaptation module composed of a bilinear upsampling layer and a convolutional layer is used to project the intermediate feature tensor of the target spatiotemporal feature extraction network in terms of spatial size and channel dimension to be consistent with the reference high-resolution feature extraction network.

[0094] Specifically, in this embodiment: to address the feature mismatch between the target network (low resolution, few channels) and the reference network (high resolution, multiple channels), we add features at the encoder ends and intermediate layers of the network. Introducing an adaptation module The mathematical expression for this module is: Its specific structure includes:

[0095] Bilinear upsampling layer: Upscales the spatial resolution of the target feature map to that of the reference feature map. Figure 1 To (e.g., scaling up from hidden layer size to) ).

[0096] Convolutional layer: Projects the number of channels of the target feature map onto the reference feature map. Figure 1 This achieves alignment of the feature space.

[0097] S32: Calculate the multi-scale composite distillation loss and the preliminary power prediction loss: Calculate the L2 norm squared distance between the adapted target feature tensor and the reference feature tensor as the semantic consistency loss; aggregate the activation value amplitudes of the feature tensors along the channel dimension to generate a spatial saliency map, and calculate the L1 norm distance between the two spatial saliency maps as the spatial saliency alignment loss; weight and sum the semantic consistency loss and the spatial saliency distribution difference loss at each level to obtain the multi-scale composite distillation loss; input the training set data into the feature extraction architecture, use the target spatiotemporal feature extraction network to output spatiotemporal features to the spatial semantic mapping and regression network, decode to generate a two-dimensional spatial influence weight map, and output the preliminary photovoltaic power prediction value through a global weighted aggregation operation; calculate the power prediction loss based on the preliminary photovoltaic power prediction value and the historical power data;

[0098] Specifically, this embodiment uses the following mathematical formula for calculation:

[0099] 1) Semantic Consistency (Feature Mimicry): This aims to make the target network mimic the feature activation patterns of the reference network. This is achieved by calculating the adapted target feature map. Reference feature map with freezing This is achieved by using the squared L2 norm distance between them:

[0100] (1)

[0101] in, Indicates the first Semantic consistency loss at each supervisory level; This indicates that the adapted target spatiotemporal feature extraction network is in the [missing information - likely a specific phase or stage]. Intermediate feature tensors of the layer; This indicates that the reference high-resolution feature extraction network is in the corresponding... The intermediate feature tensor of the layer (this feature comes from the frozen teacher network). This represents the square of the L2 norm.

[0102] 2) Spatial Saliency Alignment (Attention Mimicry): This aims to force the target network to focus on the same key regions as the reference network. First, the amplitudes of feature activations are aggregated along the channel axis to generate a spatial attention map. Then calculate the L1 norm distance between the two attention maps:

[0103] (2)

[0104] in, Indicates the first Spatial saliency alignment loss at each supervisory level; Represents the spatial attention graph of the target network; This represents the spatial attention graph of the reference network. This represents the L1 norm.

[0105] Total losses in multi-scale combined distillation: (For) The losses of each supervisory level are weighted and summed:

[0106] (3)

[0107] in, This represents the calculated total multiscale composite distillation loss; This indicates the total number of network levels involved in distillation supervision; Indicates the first Weighted balancing coefficients for hierarchical semantic consistency loss; Indicates the first Weighted balance coefficients for hierarchical spatial saliency alignment loss.

[0108] S33: Construct a joint objective function, which is composed of a weighted average of the photovoltaic power prediction loss, the cloud prediction reconstruction loss, and the multi-scale composite distillation loss. The photovoltaic power prediction loss uses mean squared error as a physical quantity-guided regularization term, generating a gradient signal for the two-dimensional spatial influence weight map during backpropagation. This gradient signal is propagated back through the global weighted aggregation layer, dynamically adjusting the weight values ​​of each pixel in the two-dimensional spatial influence weight map according to the magnitude of the prediction error. This reduces the response intensity of the target spatiotemporal feature extraction network to features in insensitive regions during the feature extraction stage. The cloud prediction reconstruction loss uses an L1 loss function to constrain the pixel-level differences between the predicted and actual cloud maps.

[0109] Specifically, in this embodiment, the total loss function is defined as:

[0110] (4)

[0111] in, To address the cloud prediction reconstruction loss, an L1 loss function is employed to ensure the visual clarity of the predicted cloud map. For photovoltaic power prediction loss, mean square error (MSE) is used as a physical constraint. To balance the hyperparameters, and based on experimental verification, the preferred setting in this embodiment is... and To balance the weights of knowledge transfer and physical regression tasks.

[0112] Parameter update strategy: Employ the AdamW optimizer, combined with a one-cycle learning rate scheduling strategy, to set the maximum learning rate (max LR) to... Weight Decay is The parameters of the target spatiotemporal feature extraction network and the spatial semantic mapping and regression network are updated synchronously using the backpropagation algorithm until the model's loss converges on the validation set.

[0113] Step S4: During the inference phase, the data path of the reference high-resolution feature extraction network is blocked, and the first resolution image sequence of the test set is input into the trained target spatiotemporal feature extraction network to infer the spatiotemporal features containing high-frequency texture information. The spatiotemporal features are then transmitted to the spatial semantic mapping and regression network to directly output the final photovoltaic power prediction value.

[0114] S41: Single-stream deployment configuration: During the model deployment phase, the reference high-resolution feature extraction network is no longer loaded;

[0115] Specifically, since the reference network (teacher model) is only used to provide supervision signals during the training phase, and it processes high-resolution ASI data ( The computational overhead is enormous. During the inference phase, the system retains only the target network (student model) with approximately 26.5M parameters, thereby reducing the computational complexity (FLOPs) to approximately 12.4G, which meets the real-time operation requirements of edge computing devices.

[0116] S42: End-to-end feature mapping and prediction: The first resolution image sequence of the test set is input into the target spatiotemporal feature extraction network, a forward convolution operation is performed to output a spatiotemporal feature tensor, and the spatiotemporal feature tensor is input into the spatial semantic mapping and regression network; the spatial semantic mapping and regression network generates the two-dimensional spatial influence weight map based on the received spatiotemporal feature tensor, and uses the learnable weight tensor to perform global weighted summation and bias correction on the map to obtain the predicted power value.

[0117] Specifically, this step includes the following two key computational sub-steps:

[0118] 1. Intermediate Representation Generation (Regression Map Generation): The spatiotemporal feature tensor output by the target spatiotemporal feature extraction network is fed into the regression network decoder. The decoder output is not directly flattened, but first passes through a final... Convolutional layers reduce their channel count to 1, thereby generating a two-dimensional spatial influence weight map (Regression Map). The spatial resolution of the weighted map is maintained at . The value of each pixel in the image. Physically representing geographical location The contribution or attenuation factor of the local sky conditions (such as whether there are clouds) to the total power of the photovoltaic power station.

[0119] 2. Global Weighted Aggregation (Power Generation): This involves generating the two-dimensional spatial influence weight map. Flattened into a vector, the final scalar photovoltaic power prediction is obtained by regression through a fully connected layer (i.e., weighted summation operation). The mathematical expression for this process is:

[0120] (5)

[0121] in, Indicates at time The predicted scalar photovoltaic power value; Represents the pixel coordinate index of the two-dimensional spatial feature map; Indicates at time The generated two-dimensional spatial influence weight map in coordinates The pixel value at the location represents the weight of the local impact of cloud cover on photovoltaic power at the corresponding geographical location. This represents a learnable global weight parameter with the same size as the weighted graph, used to weight the influence of different spatial locations; This represents the bias correction term for the regression layer.

[0122] In this way, the model achieves a direct mapping from "visual features of cloud maps" to "physical values ​​of power".

[0123] Example 2

[0124] This embodiment relates to a photovoltaic power prediction system that integrates spatial perception regression and cross-perspective knowledge transfer. The method of embodiment 1 includes: an asymmetric parallel spatiotemporal feature extraction architecture construction module, which is used to construct a target spatiotemporal feature extraction network and a reference high-resolution feature extraction network respectively, and to establish a multi-source image sequence input data stream based on geographic coordinate anchor points;

[0125] The feature dimension adaptation and multi-level cross-domain alignment execution module is used to use the reference high-resolution feature extraction network as the supervision source, and project the intermediate layer features of the target spatiotemporal feature extraction network to the reference feature space through the feature dimension adaptation and mapping unit, and calculate the multi-scale composite distillation loss including semantic consistency and spatial saliency distribution differences.

[0126] The spatial influence weight intermediate representation and power mapping generation module is used to input the spatiotemporal features output by the target spatiotemporal feature extraction network into the spatial semantic mapping and regression network, decode it to generate a two-dimensional spatial influence weight map, and output the final photovoltaic power prediction value through a global weighted aggregation operation.

[0127] The power numerical feedback-driven end-to-end collaborative calibration execution module is used to construct a joint objective function that includes spatiotemporal reconstruction error, multi-scale composite distillation error and power regression error, and uses the power numerical error as a regularization signal for collaborative parameter updates.

[0128] The single-path independent inference execution module blocks the data path of the reference high-resolution feature extraction network during the inference phase and performs photovoltaic power prediction only based on the target spatiotemporal feature extraction network.

[0129] The embodiments described in this specification are merely examples of implementations of the inventive concept. The scope of protection of this invention should not be considered as limited to the specific forms stated in the embodiments. The scope of protection of this invention also extends to equivalent technical means that can be conceived by those skilled in the art based on the inventive concept.

Claims

1. A photovoltaic power prediction method integrating spatial perception regression and cross-perspective knowledge transfer, characterized in that, Includes the following steps: S1. Collect low-frequency wide-field-of-view and high-frequency narrow-field-of-view image sequences, as well as historical power data of photovoltaic power plants, perform preprocessing, and divide them into training set, validation set, and test set; S2. Build a feature extraction architecture and a spatial semantic mapping and regression network; the feature extraction architecture includes a target spatiotemporal feature extraction network and a reference high-resolution feature extraction network; The spatial semantic mapping and regression network is connected to the output of the target spatiotemporal feature extraction network; S3. The training set data is input into the feature extraction architecture, and the spatiotemporal features output by the target spatiotemporal feature extraction network are fed to the spatial semantic mapping and regression network. After decoding, a two-dimensional spatial influence weight map is generated, and a preliminary photovoltaic power prediction value is output through a global weighted aggregation operation; the power prediction loss is calculated based on the preliminary photovoltaic power prediction value and historical power data; at the same time, using the reference high-resolution feature extraction network as the supervision source, the intermediate layer features of the target spatiotemporal feature extraction network are projected to the reference feature space through the feature dimension adaptation and mapping unit, and the package is calculated. The multi-scale composite distillation loss, incorporating semantic consistency and spatial saliency distribution differences, is used to construct a joint objective function that includes power prediction loss and multi-scale composite distillation loss. This function is then used to collaboratively update the parameters of the target spatiotemporal feature extraction network and the spatial semantic mapping and regression network. S4. During the inference phase, the data path of the reference high-resolution feature extraction network is blocked. The first-resolution image sequence of the test set is input into the trained target spatiotemporal feature extraction network to infer spatiotemporal features containing high-frequency texture information. These features are then transmitted to the spatial semantic mapping and regression network to directly output the final photovoltaic power prediction value. The calculation process of the multi-scale composite distillation loss is as follows: 1) Semantic consistency: This is achieved by calculating the adapted target feature map. Compared with reference feature map This is achieved by using the squared L2 norm distance between them: (1） in, Indicates the first Semantic consistency loss at each supervisory level; This indicates that the adapted target spatiotemporal feature extraction network is in the [missing information - likely a specific phase or stage]. Intermediate feature tensors of the layer; This indicates that the reference high-resolution feature extraction network is in the corresponding... Intermediate feature tensors of the layer; Represents the square of the L2 norm; 2) Spatial saliency alignment: First, the amplitude of feature activations is aggregated along the channel axis to generate a spatial attention map. Then calculate the L1 norm distance between the two attention maps: (2) in, Indicates the first Spatial saliency alignment loss at each supervisory level; Represents the spatial attention graph of the target network; This represents the spatial attention graph of the reference network. Represents the L1 norm; 3) Total losses from multi-scale combined distillation: For The losses from each supervisory level are weighted and summed: (3) , This represents the calculated total multiscale composite distillation loss; This indicates the total number of network levels involved in distillation supervision; Indicates the first Weighted balancing coefficients for hierarchical semantic consistency loss; Indicates the first Weighted balance coefficients for hierarchical spatial saliency alignment loss.

2. The method according to claim 1, characterized in that, Step S1 includes: S11: Data source configuration: Collect geostationary meteorological satellite cloud image data as the first resolution image sequence; collect ground-based all-sky imager data as the second resolution image sequence; collect the actual power generation of the power plant as the regression target; S12: Preprocessing and Alignment: Perform center alignment based on geographic coordinates on the first resolution image sequence and the second resolution image sequence, and perform independent statistical normalization on the image channel data and power data; S13 Data Cleaning and Dataset Division: Remove invalid samples from nighttime and those missing from sensors, and divide the processed sequence data into training set, validation set and test set.

3. The method according to claim 1, characterized in that, Step S2 describes a target spatiotemporal feature extraction network configured to process a first-resolution image sequence with low frequency and wide field of view; a reference high-resolution feature extraction network configured to process a second-resolution image sequence with high frequency and narrow field of view; and a center alignment process based on geographic coordinates is performed on the first-resolution image sequence and the second-resolution image sequence.

4. The method according to claim 3, characterized in that, The target spatiotemporal feature extraction network and the reference high-resolution feature extraction network mentioned in step S2 both adopt a hierarchical spatiotemporal prediction architecture; the architecture connects the spatial encoder, spatiotemporal converter and spatial decoder in sequence from the input end to the output end. The spatial encoder consists of multiple stacked convolutional layers, used to extract spatial features and reduce resolution; the spatiotemporal converter consists of multiple stacked Inception modules based on multi-scale parallel convolution, used to capture spatiotemporal evolution patterns while maintaining the spatial resolution of the feature map; and the spatial decoder consists of multiple sequentially stacked deconvolutional layers, used to restore spatial dimensions. The target spatiotemporal feature extraction network is configured as a lightweight architecture with fewer spatiotemporal converter layers to process image sequences at the first resolution; the reference high-resolution feature extraction network is configured as a deep architecture with more spatiotemporal converter layers to process image sequences at the second resolution. The spatial semantic mapping and regression network adopts an encoder-decoder structure with skip connections. The decoder upsamples the features and fuses the shallow spatial features from the encoder through the skip connections. At the end of the decoder, the feature map is compressed into a single-channel two-dimensional spatial influence weight map. The pixel values ​​of this map represent the local influence weight of cloud cover on photovoltaic power at the corresponding geographical location. By using a learnable weight tensor to perform weighted summation and bias correction on the two-dimensional spatial influence weight map, the scalar photovoltaic power prediction value is calculated.

5. The method as described in claim 4, characterized in that, The spatial semantic mapping and regression network specifically includes: an input adaptation layer, which uses convolutional layers to adjust the number of channels of the input features; a shrinking path, which consists of multiple convolutional blocks and downsampling layers cascaded together, used to extract deep semantic features and reduce resolution; and an expanding path, which consists of multiple upsampling layers and convolutional blocks cascaded together, where the feature map after each upsampling layer is concatenated with the feature map from the corresponding layer of the shrinking path to restore spatial resolution. Regression generation layer: Located at the end of the extension path, configured to pass through The convolutional layer compresses the multi-channel feature map into a single-channel two-dimensional spatial influence weight map, which represents the local influence weight of the sky conditions on photovoltaic power at the corresponding geographical location. The global weighted aggregation layer is configured to flatten the two-dimensional spatial influence weight map and use learnable weight parameters to perform weighted summation and bias correction on the flattened vector, directly mapping and outputting a scalar photovoltaic power prediction value.

6. The method according to claim 1, characterized in that, Step S3 includes: S31: Adapting Feature Dimensions: For multiple corresponding depth levels of the target spatiotemporal feature extraction network and the reference high-resolution feature extraction network, an adaptation module composed of a bilinear upsampling layer and a convolutional layer is used to project the intermediate feature tensor of the target spatiotemporal feature extraction network in terms of spatial size and channel dimension to be consistent with the reference high-resolution feature extraction network. S32: Calculate the multi-scale composite distillation loss and preliminary power prediction loss: Calculate the L2 norm squared distance between the adapted target feature tensor and the reference feature tensor as the semantic consistency loss; aggregate the activation value amplitudes of the feature tensors along the channel dimension to generate a spatial saliency map, and calculate the L1 norm distance between the two spatial saliency maps as the spatial saliency alignment loss; weight and sum the semantic consistency loss and spatial saliency distribution difference loss at each level to obtain the multi-scale composite distillation loss; input the training set data into the feature extraction architecture, use the target spatiotemporal feature extraction network to output spatiotemporal features to the spatial semantic mapping and regression network, decode to generate a two-dimensional spatial influence weight map, and output the preliminary photovoltaic power prediction value through a global weighted aggregation operation; calculate the power prediction loss based on the preliminary photovoltaic power prediction value and the historical power data; S33: Construct a joint objective function, which is composed of a weighted average of photovoltaic power prediction loss, cloud prediction reconstruction loss, and the multi-scale composite distillation loss; the power prediction loss uses mean square error and is configured as a physical quantity-guided regularization term; during the collaborative update of parameters, the gradient signal generated by the power prediction loss is used to back-calibrate the response weights of the target spatiotemporal feature extraction network to the input features; while the cloud prediction reconstruction loss uses an L1 loss function to constrain the pixel-level difference between the predicted cloud map and the real cloud map.

7. The method according to claim 1, characterized in that, Step S4 includes: S41: Single-stream deployment configuration: During the model deployment phase, the reference high-resolution feature extraction network is no longer loaded; S42: End-to-end feature mapping and prediction: The first resolution image sequence of the test set is input into the target spatiotemporal feature extraction network, a forward convolution operation is performed to output a spatiotemporal feature tensor, and the spatiotemporal feature tensor is input into the spatial semantic mapping and regression network; the spatial semantic mapping and regression network generates the two-dimensional spatial influence weight map based on the received spatiotemporal feature tensor, and uses the learnable weight tensor to perform global weighted summation and bias correction on the map to obtain the predicted power value.

8. The method as described in claim 7, characterized in that, Step S42, the global weighted aggregation, includes: [the process of] the weight graph... Flatten the image and regress it through a fully connected layer to obtain the final scalar photovoltaic power prediction.

9. A photovoltaic power prediction system integrating spatial perception regression and cross-perspective knowledge transfer, characterized in that: include: An asymmetric parallel spatiotemporal feature extraction architecture building module is used to construct the target spatiotemporal feature extraction network and the reference high-resolution feature extraction network respectively, and to establish a multi-source image sequence input data stream based on geographic coordinate anchor points; The feature dimension adaptation and multi-level cross-domain alignment execution module is used to project the intermediate layer features of the target spatiotemporal feature extraction network onto the reference feature space through the feature dimension adaptation and mapping unit, using the reference high-resolution feature extraction network as the supervision source, and to calculate the multi-scale composite distillation loss, which includes semantic consistency and spatial saliency distribution differences; the calculation process of the multi-scale composite distillation loss is as follows: 1) Semantic consistency: This is achieved by calculating the adapted target feature map. Compared with reference feature map This is achieved by using the squared L2 norm distance between them: (1） in, Indicates the first Semantic consistency loss at each supervisory level; This indicates that the adapted target spatiotemporal feature extraction network is in the [missing information - likely a specific phase or stage]. Intermediate feature tensors of the layer; This indicates that the reference high-resolution feature extraction network is in the corresponding... Intermediate feature tensors of the layer; Represents the square of the L2 norm; 2) Spatial saliency alignment: First, the amplitude of feature activations is aggregated along the channel axis to generate a spatial attention map. Then calculate the L1 norm distance between the two attention maps: (2) in, Indicates the first Spatial saliency alignment loss at each supervisory level; Represents the spatial attention graph of the target network; This represents the spatial attention graph of the reference network. Represents the L1 norm; 3) Total losses from multi-scale combined distillation: For The losses from each supervisory level are weighted and summed: (3) , This represents the calculated total multiscale composite distillation loss; This indicates the total number of network levels involved in distillation supervision; Indicates the first Weighted balancing coefficients for hierarchical semantic consistency loss; Indicates the first Weighted balance coefficients for hierarchical spatial saliency alignment loss; The spatial influence weight intermediate representation and power mapping generation module is used to input the spatiotemporal features output by the target spatiotemporal feature extraction network into the spatial semantic mapping and regression network, generate a two-dimensional spatial influence weight map after decoding, and output the photovoltaic power prediction value through global weighted aggregation operation; The power numerical feedback-driven end-to-end collaborative calibration execution module is used to construct a joint objective function that includes spatiotemporal reconstruction error, multi-scale composite distillation error and power regression error, and uses the power numerical error as a regularization signal for collaborative parameter updates. The single-path independent inference execution module blocks the data path of the reference high-resolution feature extraction network during the inference phase and performs photovoltaic power prediction only based on the target spatiotemporal feature extraction network.

Images