Methods, products and equipment for photovoltaic power forecasting based on space-ground collaborative observation

Abstract

This invention provides a method, product, and equipment for predicting the power of photovoltaic (PV) power plants using satellite-ground collaborative observation. The method includes: acquiring satellite-based and ground-based observation data of the PV power plant's location; constructing engineering-derived features based on the ground-based observation data, and fusing the satellite-based, ground-based, and engineering-derived features into multi-source observation features; processing the multi-source observation features using multiple pre-trained deep heterogeneous basis models to obtain multiple depth prediction sequences; performing daytime masking and two-layer signal decomposition on historical available power sequences, and obtaining prediction tensors for multiple decomposition channels through analytical extrapolation; inputting the depth prediction sequences and prediction tensors into an attention-integrated decision module, which then outputs the final power prediction result. This method enables complementary utilization of satellite-ground collaborative observation information, while reducing prediction costs and improving the accuracy of power prediction.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic technology, and in particular to a method, product, and equipment for forecasting the power of photovoltaic power plants using satellite-ground collaborative observation. Background Technology

[0002] Currently, photovoltaic (PV) power generation accounts for an increasingly larger share of my country's power system, becoming the mainstay of new energy installations. However, PV output is highly random and volatile, influenced by meteorological factors such as solar irradiance, cloud cover, and temperature, posing significant challenges to grid dispatching, reserve capacity allocation, and new energy consumption. Accurate short-term forecasting of available power from PV power plants is a key technological foundation for ensuring the safe and stable operation of the power system, improving the new energy consumption rate, and reducing system reserve capacity requirements. Summary of the Invention

[0003] In view of the above problems, the present invention proposes a method, product and equipment for predicting the power of photovoltaic power plants by space-ground collaborative observation to overcome the above problems or at least partially solve the above problems.

[0004] One objective of this invention is to improve the accuracy of power prediction for photovoltaic power plants and reduce costs.

[0005] Specifically, this invention provides a method for predicting the power of photovoltaic power plants based on satellite-ground collaborative observation, comprising: Acquire satellite-based and ground-based observation data of the location of the photovoltaic power station; Engineering-derived features are constructed based on ground-based observation data, and satellite-based observation data, ground-based observation data, and engineering-derived features are integrated into multi-source observation features. Multiple pre-trained deep heterogeneous basis models are used to process multi-source observation features to obtain multiple deep prediction sequences. The deep prediction sequences contain power prediction values ​​for multiple future time steps. The deep heterogeneous basis models are used to extract temporal dependencies from multi-source observation features and map them to obtain power prediction values ​​for multiple future time steps. Daytime masking and two-layer signal decomposition are performed on the historical available power sequence, and the prediction tensors of multiple decomposition channels are obtained by analytical extrapolation. The deep prediction sequence and prediction tensor are input into the attention fusion decision module, which outputs the final power prediction result. The attention fusion decision module is used to fuse the deep prediction sequence and prediction tensor through the attention mechanism to generate the final power prediction result.

[0006] Optionally, the steps of performing daytime masking and two-layer signal decomposition on the historical available power sequence, and obtaining the prediction tensor of the decomposed channel by analytical extrapolation, include: Construct daytime masking rules, and filter out available power sequences in the preset daytime period from the historical available power sequences according to the daytime masking rules; The first-level decomposition is performed on the available power sequence to obtain seasonal periodic components, trend components and residual components. The first-level decomposition includes the seasonal trend decomposition method. By performing linear extrapolation on the trend component, the predicted values ​​of the trend component at multiple future time steps are obtained; A daily cycle template is constructed based on the seasonal cycle components; Based on the daily cycle template, the seasonal cycle components are extrapolated by cycle reuse to obtain the predicted values ​​of the seasonal cycle components in the future multiple time steps. A second-level decomposition is performed on the residual components to obtain multiple intrinsic mode components. The second-level decomposition includes variational mode decomposition. Extrapolate each intrinsic mode component using a first-order autoregressive model to obtain the predicted value of each intrinsic mode component at multiple future time steps. The predicted values ​​of the extrapolated seasonal cycle component, the predicted value of the trend component, and the predicted values ​​of multiple intrinsic mode components are summed to obtain the prediction tensor of the full component reconstruction summation channel. The predicted values ​​of the seasonal cycle component, the trend component, and each intrinsic mode component are used as the predicted values ​​of the seasonal channel, the trend channel, and each intrinsic mode component channel, respectively, thus obtaining the predicted tensors of multiple decomposed channels.

[0007] Optionally, the intrinsic model components obtained from the variational model decomposition include hourly weather processes, 10- to 30-minute cloud movement processes, minute-level fluctuations, and high-frequency noise. The decomposition channels include seasonal channels, trend channels, channels for each intrinsic mode component, and full component reconstruction summation channels. The seasonal channel corresponds to the predicted value of the seasonal cycle component at multiple future time steps, the trend channel corresponds to the predicted value of the trend component at multiple future time steps, and the predicted value of the full component reconstruction summation channel is obtained by summing the predicted values ​​of the seasonal cycle component, the trend component, and each intrinsic mode component.

[0008] Optionally, the calculation formula for periodic extrapolation of seasonal periodic components is as follows: Where t represents the current time and h represents the sequence number of the future time step. This represents the predicted value of the seasonal cycle component at the h-th step in the future. This represents a daily periodic template function, whose input is the position number of the time step within a day. Indicates the position number at the h-th step in the future; The formula for linear extrapolating the trend component is as follows: in, This represents the predicted value of the trend component at the h-th time step in the future, where a and b are the slope and intercept of the linear extrapolation, respectively. The trend component values ​​at the last multiple time steps are fitted using the least squares method. The formula for extrapolating each intrinsic mode component using a first-order autoregressive model is as follows: in, This represents the predicted value of the k-th intrinsic mode component at the h-th step in the future. The first-order autoregressive coefficient of the k-th intrinsic mode component is estimated from the samples of the last few time steps of this component. Let K be the average of the samples of the k-th intrinsic mode component over the last multiple time steps, where K is the number of intrinsic mode components. The formula for calculating the prediction tensor of the full component reconstruction summation channel is: in, This represents the prediction tensor of the full component reconstruction summation channel.

[0009] Optionally, the steps of inputting the depth prediction sequence and prediction tensor into the attention ensemble decision module, and having the attention ensemble decision module output the final power prediction result, include: The deep prediction sequence and prediction tensor are combined as the input tensor of the attention-integrated decision module; In the static prior branch of the attention integration decision module, multiple deep prediction sequences are obtained, and static prior weights are obtained on the validation set through mean square error grid search. The static prior weights are then weighted and aggregated to output a stable baseline. In the dynamic attention branch of the attention integration decision module, features are constructed based on the input tensor, and the feature construction results are passed through a linear layer, layer normalization, modified linear unit, and dropout method to obtain the hidden representation. The output of the linear layer has a logical value with the same dimension as the total number of base predictors. The attention weight is obtained by normalizing through the normalized exponential function. The attention weight is used to sum the predicted values ​​of each base predictor at this time step to obtain the dynamic weighted prediction result. In the residual correction branch of the attention integration decision module, the hidden representation of the dynamic attention branch is reused, and residual gating coefficients are generated through linear layers and preset activation functions, respectively. Additive residuals are generated through linear layers and modified linear units. The stable baseline, dynamic weighted prediction results, residual gating coefficients, and additive residuals are calculated according to a preset formula to obtain the final power prediction results.

[0010] Optionally, the formula for calculating the final power prediction result is as follows: in, Let M represent the final power prediction result at time step h, where M represents the number of deep heterogeneous basis models and C represents the number of prediction tensors. This represents the power prediction value of the m-th depth prediction sequence at the h-th time step. This represents the power prediction value of the c-th base predictor at time step h. The base predictor includes a depth prediction sequence and a prediction tensor. This represents the static prior weights of the m-th depth prediction sequence. This represents the dynamic attention weights of the c-th base predictor at the h-th time step. This represents the residual gating coefficient at the h-th time step. This represents the additive residual at the h-th time step.

[0011] Optionally, the satellite-based observation data includes albedo data and brightness temperature data for multiple bands from geostationary meteorological satellites; Ground-based observation data includes total radiation data measured by the site's radiometer and multi-channel measured power data from the field. Engineering-derived features include one or more of the following features: periodic coding features, aggregate power features, multi-scale lag features, ultra-short-term rate of change features, rolling statistical features, and efficiency ratio features.

[0012] Optionally, the satellite-end observation data includes albedo data in 6 visible or near-infrared bands and brightness temperature data in 10 infrared bands, for a total of 16 dimensions of data; Ground-based observation data includes 1-dimensional total radiation data and 4-channel measured power data, for a total of 5 dimensions; The engineering-derived features specifically include: 6-dimensional periodic coding features, 1-dimensional aggregated power features, 5-dimensional multi-scale lag features, 3-dimensional ultra-short-term change rate features, 2-dimensional rolling statistical features, and 1-dimensional efficiency ratio features, for a total of 19 features.

[0013] According to another aspect of the present invention, a computer program product is also provided, comprising a computer program that, when executed by a processor, implements the steps of the above-described method for predicting photovoltaic power generation from satellite-ground co-observation.

[0014] According to another aspect of the present invention, a computer device is also provided, including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor computer program is used to implement the above-described method for predicting the power of photovoltaic power plants through satellite-ground collaborative observation.

[0015] The photovoltaic power prediction method based on satellite-ground collaborative observation of the present invention first acquires satellite-end observation data and ground-end observation data of the location of the photovoltaic power station; then, it constructs engineering-derived features based on the ground-end observation data and fuses the satellite-end observation data, ground-end observation data, and engineering-derived features into multi-source observation features; next, it processes the multi-source observation features using multiple pre-trained deep heterogeneous basis models to obtain multiple deep prediction sequences, which contain power prediction values ​​for multiple future time steps. The deep heterogeneous basis models are used to extract temporal dependencies from the multi-source observation features and map them to obtain power prediction values ​​for multiple future time steps; daytime masking and two-layer signal decomposition are performed on the historical available power sequences, and prediction tensors of multiple decomposition channels are obtained through analytical extrapolation; finally, the deep prediction sequences and prediction tensors are input into the attention integration decision module, which outputs the final power prediction result. The attention integration decision module is used to fuse the deep prediction sequences and prediction tensors through an attention mechanism to generate the final power prediction result. This method enables the complementary use of satellite-ground collaborative observation information, avoids the contamination of zero-value pairs at night, greatly reduces the cost of forecasting, and improves the accuracy of photovoltaic power forecasting.

[0016] The above and other objects, advantages and features of the present invention will become more apparent to those skilled in the art from the following detailed description of specific embodiments of the invention in conjunction with the accompanying drawings. Attached Figure Description

[0017] The following sections will describe some specific embodiments of the invention in detail by way of example and not limitation, with reference to the accompanying drawings. The same reference numerals in the drawings denote the same or similar parts or portions. Those skilled in the art should understand that these drawings are not necessarily drawn to scale. In the drawings: Figure 1 This is a schematic diagram of the architecture of a photovoltaic power prediction method based on satellite-ground collaborative observation according to an embodiment of the present invention; Figure 2 This is a flowchart illustrating a photovoltaic power forecasting method based on satellite-ground collaborative observation according to an embodiment of the present invention. Figure 3 This is a schematic diagram of the process of daytime mask screening, double-layer signal decomposition, and analytical extrapolation in a photovoltaic power prediction method based on satellite-ground collaborative observation according to an embodiment of the present invention. Figure 4 This is a schematic diagram of the execution flow of the attention integration decision module in a photovoltaic power forecasting method based on satellite-ground collaborative observation according to an embodiment of the present invention. Figure 5 This is an average attention weight heatmap of a photovoltaic power prediction method based on satellite-ground collaborative observation according to an embodiment of the present invention; Figure 6 This is a scatter plot of the photovoltaic power forecast for the next 4 hours in the satellite-ground coordinated observation photovoltaic power forecasting method according to an embodiment of the present invention; Figure 7 This is a scatter plot of the photovoltaic power forecast for the 30th minute in the future, generated by satellite-ground collaborative observation in a photovoltaic power forecasting method according to an embodiment of the present invention. Figure 8 This is a schematic diagram of photovoltaic power forecasting using satellite-ground coordinated observation in a photovoltaic power forecasting method according to another embodiment of the present invention; Figure 9 This is a schematic diagram of a computer program product according to an embodiment of the present invention; Figure 10 This is a schematic diagram of a computer-readable storage medium according to an embodiment of the present invention; and Figure 11 A schematic diagram of a computer device according to an embodiment of the present invention. Detailed Implementation

[0018] Those skilled in the art should understand that the embodiments described below are merely a part of the embodiments of the present invention, and not all of the embodiments of the present invention. These partial embodiments are intended to explain the technical principles of the present invention and are not intended to limit the scope of protection of the present invention. Based on the embodiments provided by the present invention, all other embodiments obtained by those skilled in the art without creative effort should still fall within the scope of protection of the present invention.

[0019] It should be noted that the logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be specifically implemented in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus or device (such as a computer-based system, a processor-included system or other system that can fetch and execute instructions from, an instruction execution system, apparatus or device).

[0020] In engineering, ultra-short-term forecasting generally refers to rolling forecasts of available power within the next 0-4 hours. Power dispatching pays particular attention to the forecast accuracy for the period 15-30 minutes in advance (as it relates to the issuance of AGC commands), and the performance indicator uses the "short-term cumulative absolute error ratio" rather than the traditional mean square error. This means that the optimization objective of the forecasting algorithm must be explicitly aligned with the performance indicator.

[0021] Figure 1 This is a schematic diagram of the architecture of a photovoltaic power forecasting method based on satellite-ground collaborative observation according to an embodiment of the present invention, as shown below. Figure 1As shown, the architecture includes a data source layer + feature engineering 110, a deep model layer + decomposition layer 120, and an integrated decision layer 130. The data source layer + feature engineering 110 further includes a satellite-meteorological satellite 111 and a ground-based photovoltaic power station 112, and generates multi-source observation features through a satellite-ground collaborative multi-source observation feature construction module 113. The deep model layer + decomposition layer 120 includes a multi-regularized deep model training module 121, four deep heterogeneous base models (BiLSTM-Attention 122, CNN-LSTM 123, Transformer 124, TCN 125), and a daytime mask-based STL+VMD two-layer signal decomposition and analytical extrapolation module 126. The integrated decision layer 130 includes a residual-corrected attention stacking integrated decision module 131 and outputs the final power prediction result.

[0022] Based on this architecture, the method of this invention first acquires satellite-end observation data and ground-end observation data in the data source layer + feature engineering 110. Satellite-end observation data refers to albedo and brightness temperature data from multiple bands of geostationary meteorological satellites (such as Himawari-9), which can reflect the spatial distribution and movement trend of clouds in the sky; ground-end observation data refers to the total radiation measured by the radiometer and multiple measured power data from the photovoltaic power station, which can accurately characterize the current power response state of the station. Through the satellite-ground collaborative multi-source observation feature construction module 113, the original satellite-end and ground-end observation data are fused with engineering-derived features (such as periodic coding, aggregated power, multi-scale lag, ultra-short-term change rate, rolling statistics, efficiency ratio, etc.) to form multi-source observation features.

[0023] In some optional embodiments, the satellite-end observation data generally includes 16-dimensional data, as shown in Table 1: Table 1 Satellite-end observation data Statistical analysis showed that the correlation coefficients between albedo_04, albedo_05 and available power (kygl) were r = 0.58-0.59; for tbb_07, r ≈ 0.51; and for the other satellite channels, r = 0.02-0.06. However, these inputs were retained as supplementary information for multi-source fusion.

[0024] Ground-based observation data typically includes 5-dimensional data, as shown in Table 2: Table 2 Ground-level observation data Engineering-derived features typically include 19 dimensions of data, providing short-term dynamic information based on ground-end target variables and power plant power structure. Specific examples are shown in Table 3. Table 3 Engineering Derivative Characteristics Thus, the final input feature dimension is 16+5+19=40 dimensions. Satellite observation (satellite end) reflects the distribution of clouds / radiation fields in the sky, with the advantage of spatial dimension, and can predict the arrival of clouds 30~120 minutes later; ground observation (ground end) reflects the current actual response of the station, with the advantages of temporal dimension and local state, and can accurately characterize the power dynamics within the past 10 minutes; engineering-derived features act as a bridge between the two ends, encoding the current measured response and historical dynamic trends into the same time window.

[0025] The reason for explicitly separating satellite-end and ground-end observations and superimposing engineering features is that the accuracy of a single observation source drops sharply under cloudy or abrupt weather conditions: satellite observations can sense cloud movement but cannot reflect the instantaneous response of the site, while ground observations, although accurate, lack spatial information. The "satellite-ground-time" triangular coverage formed by these three sources allows the other two ends to compensate when any one source is blocked or abnormal, thereby significantly improving the robustness of predictions.

[0026] In the deep model layer + decomposition layer 120, on the one hand, the multi-regularized deep model training module 121 independently trains four deep heterogeneous base models under a unified multi-regularization framework (e.g., AdamW weight decay, high-ratio dropout, layer normalization, and Gaussian noise injection during training): a Bidirectional Long Short-Term Memory Network (BiLSTM-Attention) 122, a Convolutional Long Short-Term Memory Network (CNN-LSTM) 123, a Transformer 124, and a Temporal Convolutional Network (TCN) 125. These models extract temporal dependencies from multi-source observation features and output multiple deep prediction sequences. Each sequence typically contains power predictions for the next 24 time steps (corresponding to 4 hours at a resolution of 10 minutes). The reason for using heterogeneous model ensemble is that different models have different inductive biases, which can capture patterns in the data from different perspectives and reduce the risk of overfitting by a single model.

[0027] On the other hand, the STL+VMD dual-layer signal decomposition and analytical extrapolation module 126 under the daytime mask processes the historical available power sequence: First, a daytime mask rule is constructed (for example, only the daytime period from 06:30 to 19:00 is retained), and about 50.9% of the zero values ​​at night are removed; then, the first layer of STL decomposition (cycle 144, smoothing parameters 7 and 1441) is performed on the daytime period sequence to obtain the seasonal cycle, trend and residual components; then, a daily cycle template of length 144 is constructed based on the seasonal cycle component; then, the second layer of VMD decomposition (K=4) is performed on the residual component to obtain four intrinsic mode components, which correspond to hourly weather processes, 10-30 minute cloud movement, minute-level fluctuations and high-frequency noise, respectively; finally, through three analytical extrapolation methods, cycle reuse, linear extrapolation (last 30 steps least squares fitting) and AR(1) recursion (last 12 steps coefficient estimation), the prediction tensor of 7 decomposition channels (seasonal channel, trend channel, four IMF channels and full component reconstruction summation channel) is generated. The reason for using a daytime mask is to avoid the negative oscillations caused by zero values ​​at night, which would contaminate the decomposition results and give the frequency bands learned by the downstream model a clear physical meaning. Instead of training a deep model separately for each IMF, analytical extrapolation is used, so that the overall training time is independent of the number of IMFs. Compared with the existing route, the training time is reduced from 16-20 hours to about 30 minutes, which greatly reduces the computational cost.

[0028] In the integrated decision layer 130, the residual-corrected attention stacking integrated decision module 131 combines the prediction tensors of 4 deep prediction sequences and 7 decomposed channels (a total of 11 base predictors) into an input tensor. This module contains three branches: the static prior branch determines fixed weights for only the 4 deep prediction sequences on the validation set through mean squared error grid search, outputting a stable baseline; the dynamic attention branch extracts the predicted values ​​and their statistical characteristics (mean, standard deviation, range, etc.) of each base predictor from the input tensor for each prediction time step, and combines them with step-level position encoding, generates a hidden representation through a multilayer perceptron (linear layer-layer normalization-corrected linear unit-dropout method), then obtains attention weights through a linear layer and a normalized exponential function, and finally weights and sums the results of each base predictor to obtain a dynamically weighted prediction; the residual correction branch reuses the hidden representation, generates residual gating coefficients through a linear layer and a sigmoid function, and generates additive residuals through a linear layer and a corrected linear unit.

[0029] The reason for using this three-branch architecture is that: static priors provide a stable baseline, ensuring that initial performance is no worse than pure deep learning model ensembles; dynamic attention enables adaptive weight allocation at the sample and step levels, compensating for the shortcomings of fixed weights; residual gating limits the deviation of dynamic corrections, avoiding overfitting. The synergy of these three elements makes the ensemble results both stable and adaptive.

[0030] The formula for calculating the final power prediction result is shown in equation (1): Equation (1) in, This represents the final power prediction result at the s-th prediction time step. Indicates a stable baseline. This indicates a dynamically weighted prediction. Represents the residual gating coefficient. This represents additive residuals.

[0031] By working together, the above modules can achieve efficient and high-precision ultra-short-term power forecasting for photovoltaic power plants, and can directly serve the grid dispatching's assessment requirements for 30-minute advance forecasting.

[0032] Figure 2 This is a flowchart illustrating a photovoltaic power forecasting method based on satellite-ground collaborative observation according to an embodiment of the present invention, as shown below. Figure 2 As shown, the process includes at least the following steps S201 to S205.

[0033] Step S201: Acquire satellite-based and ground-based observation data of the photovoltaic power station's location. Satellite-based observation data typically includes albedo data and brightness temperature data from multiple bands of geostationary meteorological satellites. In a preferred embodiment, the satellite-based observation data may include albedo data for six visible / near-infrared bands (albedo_01~albedo_06) and brightness temperature data for ten infrared bands (tbb_07~tbb_16), totaling 16 dimensions. This data has spatial distribution sensing capabilities and can predict cloud arrival 30–120 minutes in advance. Ground-based observation data comes from within the photovoltaic power station and meteorological stations, including total radiation measured by the station's radiometer and multiple measured power outputs (e.g., p1, p2, p3, p4). In a preferred embodiment, the ground-based observation data typically includes one dimension of total radiation and four measured power outputs, totaling five dimensions. The correlation coefficient between total radiation and available power is as high as 0.974, and the correlation coefficient between measured power and available power is greater than 0.99, which can accurately characterize the power dynamic response of the station within the past 10 minutes.

[0034] This step, by simultaneously acquiring data from both satellite and ground sources, provides a physically complementary information foundation for subsequent fusion, and solves the problem of a sharp decline in accuracy of a single observation source (satellite only or ground only) under cloudy or sudden weather conditions.

[0035] Step S202 involves constructing engineering-derived features based on ground-based observation data and fusing satellite-based observation data, ground-based observation data, and engineering-derived features into multi-source observation features. After acquiring the raw data, engineering-derived features are further constructed using ground-based data. These features aim to enhance the model's ability to capture short-term dynamics and periodic patterns. Specifically, the engineering-derived features include one or more of the following: periodic encoding features (such as sine / cosine encoding for hours, months, and days), aggregated power features (average of four power sources), multi-scale lag features (target value displacement of 1, 2, 3, 6, and 12 steps), ultra-short-term rate of change features (first-order difference), rolling statistics features (mean and standard deviation of the last hour), and efficiency ratio features (ratio of available power to measured power). In a preferred embodiment, a total of 19 engineering-derived features are constructed: 6-dimensional periodic encoding, 1-dimensional aggregated power, 5-dimensional multi-scale lag, 3-dimensional ultra-short-term rate of change, 2-dimensional rolling statistics, and 1-dimensional efficiency ratio. Subsequently, the satellite-based observation data (16-dimensional), ground-based observation data (5-dimensional), and the aforementioned engineering-derived features (19-dimensional) are stitched together to form a 40-dimensional multi-source observation feature. This fused feature retains the spatial advantages of satellite observations while incorporating the precise response of ground-based observations and prior engineering knowledge, forming a "satellite-ground-time" triangular coverage. When any source fails due to obstruction or anomalies, the information from the other two sources can compensate for each other, thereby significantly improving the robustness of predictions.

[0036] Step S203: Multiple pre-trained deep heterogeneous basis models are used to process the multi-source observation features to obtain multiple deep prediction sequences. These deep prediction sequences typically contain power predictions for multiple future time steps. Deep heterogeneous basis models are generally used to extract temporal dependencies from multi-source observation features and map them to obtain power predictions for multiple future time steps.

[0037] The fused multi-source observation features are input into multiple pre-trained deep heterogeneous base models. These models are used to extract temporal dependencies from the multi-source observation features and map them to obtain power predictions for multiple future time steps. In this embodiment, there are four deep heterogeneous base models: a Bidirectional Long Short-Term Memory Network (BiLSTM-Attention), a Convolutional Long Short-Term Memory Network (CNN-LSTM), a Transformer, and a Temporal Convolutional Network (TCN). Each model is trained independently and outputs power predictions for the next 24 time steps (with a resolution of 10 minutes, i.e., the next 4 hours). These values ​​constitute multiple deep prediction sequences. The purpose of using heterogeneous model ensemble is to leverage the inductive biases of different models to reduce the risk of overfitting from a single model, thereby improving the stability and generalization ability of the predictions.

[0038] Optionally, during the training phase, a unified multi-regularization framework can generally be used to prevent overfitting, including AdamW weight decay (weight_decay=1×10). -4 The methods used include a uniform high-ratio dropout method (Dropout=0.35) across the entire model, layer normalization (LayerNorm), and adding Gaussian noise (N(0,0.02)) to the training data. 2 These regularization techniques are applied to the parameter space, network structure, feature distribution, and data layer, respectively, and can synergistically improve the generalization performance of the model.

[0039] Step S204 involves performing daytime masking and two-layer signal decomposition on the historical available power sequence, and obtaining the prediction tensors of multiple decomposition channels through analytical extrapolation. This step processes the target variable—the historical available power sequence (i.e., the time series of effective power output by the power grid hub nodes)—to obtain virtual basis predictions that do not rely on neural networks.

[0040] In some alternative embodiments, a daytime masking rule is first constructed, for example, M={t|06:30≤time_of_day(t)<19:00}, which retains only the historical available power data during the daytime period and removes about 50.9% of the zero values ​​at night to avoid the generation of pseudo-oscillating intrinsic mode components by the zero values ​​at night during decomposition.

[0041] Then, a two-level signal decomposition is performed on the selected daytime time series: the first level of decomposition adopts a seasonal-trend decomposition method based on local weighted regression. The parameters can generally be set as follows: seasonal period length period=144 (corresponding to 24 hours × 6 steps / hour), seasonal smoother seasonal_smoother=7, trend smoother trend_smoother=1441, robust=True. The decomposition yields the seasonal period component S(t), trend component T(t), and residual component R(t). Next, a daily period template of length 144 is constructed based on the seasonal period component. The value at each position in the template is taken as the median of the seasonal components at all the same positions in the training set.

[0042] The second-level decomposition applies Variational Mode Decomposition (VMD) to the residual component R(t), setting the number of decompositions K=4, the penalty parameter α=2000, and τ=0. After solving the constrained variational problem, four intrinsic mode components (IMF1~IMF4) are obtained, whose frequency bands, from low to high, correspond to hourly weather processes, 10-30 minute cloud movement processes, minute-level fluctuations, and high-frequency noise, respectively.

[0043] After decomposition, analytical extrapolation is used (without training any neural network) to obtain the predicted values ​​for the next 24 steps: Specifically, for seasonal periodic components, periodic reuse can generally be used, and the calculation formula is shown in equation (2): Equation (2) Where t represents the current time and h represents the sequence number of the future time step. This represents the predicted value of the seasonal cycle component at the h-th step in the future. This represents a daily periodic template function, whose input is the position number of the time step within a day. This indicates the position number of the h-th step in the future, meaning that the table can be looked up directly based on the position number of the future time step within a day.

[0044] For trend components, linear extrapolation can generally be used, and the calculation formula is shown in equation (3): Equation (3) in, Let represent the predicted value of the trend component at the h-th time step in the future, where a and b are the slope and intercept of the linear extrapolation, respectively, and are obtained by fitting the trend component values ​​of the last multiple time steps using the least squares method.

[0045] For each intrinsic mode component, a first-order autoregressive model is used for recursion, and the calculation formula is shown in equation (4): Equation (4) in, This represents the predicted value of the k-th intrinsic mode component at the h-th step in the future. The first-order autoregressive coefficient of the k-th intrinsic mode component is estimated from the samples of the last few time steps of this component. For example, this component can be estimated from the samples of the last 12 time steps. Let K be the sample mean of the k-th intrinsic mode component over the last multiple time steps, where K is the number of intrinsic mode components.

[0046] The calculation formula for the prediction tensor of the full component reconstruction summation channel is shown in equation (5): Equation (5) in, This represents the prediction tensor of the full component reconstruction summation channel.

[0047] Finally, the extrapolation results are combined into prediction tensors for 7 decomposed channels: seasonal channel (stl_seasonal), trend channel (stl_trend), four IMF channels (vmd_imf_1~vmd_imf_4), and full component reconstruction summation channel (decomp_sum, i.e., the sum of seasonal + trend + four IMFs).

[0048] These channels serve as "virtual basis predictions" input to the subsequent ensemble decision layer. By using daytime masking, the contamination of the decomposition by nighttime zero values ​​is avoided, giving each intrinsic mode component a clear physical meaning. By employing analytical extrapolation instead of training a deep model separately for each IMF, the overall training workload becomes independent of the number of IMFs. Compared to the traditional "decomposition + serial retraining" approach, the training time is reduced from 16-20 hours to approximately 30 minutes, significantly saving computational costs.

[0049] Step S205: The deep prediction sequence and prediction tensor are input into the attention ensemble decision module, and the attention ensemble decision model outputs the final power prediction result. The attention ensemble decision module is generally used to fuse the deep prediction sequence and prediction tensor through an attention mechanism to generate the final power prediction result.

[0050] Based on the above embodiments, this step combines the four depth prediction sequences obtained in step S203 and the seven decomposed channel prediction tensors (a total of 11 base predictors) obtained in step S204 into an input tensor, which is then fed into the attention ensemble decision module. This module adopts a residual-corrected attention stacker generalization architecture, which generally includes three branches: Static prior branch: For only the four depth prediction sequences, a set of fixed weights (with the weights of the decomposed channels always being 0) is determined on the validation set through mean squared error grid search, and the weighted average is calculated as a stable baseline. This baseline ensures that the ensemble results are at least no worse than the pure depth model ensemble.

[0051] The dynamic attention branch extracts the predicted values ​​and their statistical features (mean, standard deviation, maximum, minimum, range, etc.) of all base predictors at each prediction time step h from the input tensor. This is then concatenated with step-level positional encoding and static prior mixed features, and sequentially passed through a linear layer, layer normalization, Corrected Linear Unit (ReLU), and dropout method to obtain the hidden representation. The output of the linear layer is then converted into 11-dimensional logistic values ​​via a normalized exponential function (softmax) to form attention weights. Finally, the predicted values ​​of each base predictor are weighted and summed to obtain a dynamically weighted prediction. This branch implements adaptive weight allocation at both the sample and step levels, dynamically adjusting the importance of each base predictor based on different weather conditions and prediction distances.

[0052] Residual correction branch: The hidden representation of the dynamic attention branch is reused to generate residual gating coefficients through linear layers and sigmoid functions, and additive residuals through linear layers and ReLU. The gating coefficients control the deviation of the dynamic prediction from the static baseline. Initially, dynamic correction is suppressed, and the optimal value is gradually learned during training to avoid overfitting. Then, the results are fused according to Equation (1) to obtain the final prediction result.

[0053] This method achieves several key benefits. First, the construction of multi-source observation features through satellite-ground collaboration integrates the spatial coverage advantages of satellite remote sensing with the local accuracy of ground-based measurements, forming a "satellite-ground-time" triangular coverage. This effectively solves the problem of a sharp decline in accuracy from a single observation source under cloudy or abrupt weather conditions, directly meeting the high requirements of power grid dispatch for short-term forecasting. Second, the STL+VMD two-layer decomposition and periodic reuse, linear extrapolation, and first-order autoregressive analytical extrapolation strategy under daytime masking not only eliminates approximately 50.9% of the zero values ​​at night that contaminate the decomposition, giving the intrinsic mode components a clear physical meaning, but also eliminates the need to train deep models for each component separately. This makes the overall training workload independent of the number of components, reducing the training time from 16-20 hours in the traditional "decomposition + serial retraining" route to approximately 30 minutes, saving about 30 times the computational cost. Third, the residual-corrected attention-based integrated decision module, through the fusion of static priors, dynamic attention, and residual gating, possesses both stability and adaptability, meeting the requirements of real-time dispatching. Meanwhile, this method has good interpretability, maintainability and portability, and can be flexibly adapted to different satellite models and field station data interfaces, which facilitates engineering deployment and iterative updates.

[0054] Figure 3 This is a schematic diagram illustrating the process of daytime mask screening, two-layer signal decomposition, and analytical extrapolation in a photovoltaic power forecasting method based on satellite-ground coordinated observation according to an embodiment of the present invention. Figure 3 As shown, the process includes at least the following steps S301 to S309.

[0055] Step S301: Construct daytime masking rules, and filter out available power sequences in the preset daytime period from the historical available power sequences according to the daytime masking rules.

[0056] First, a daytime masking rule is constructed, for example, M = {t | 06:30 ≤ time_of_day(t) < 19:00}. Based on this rule, power data within a preset daytime period (e.g., 06:30 to 19:00) is extracted from the historical available power sequence to obtain the daytime available power sequence. The reason for this daytime masking is that photovoltaic power is almost zero at night (approximately 50.9% of the total sample). If the entire 24-hour sequence were directly decomposed, a large number of zero values ​​would generate spurious oscillations in the intrinsic mode components, rendering the decomposition results meaningless. By retaining only daytime data, the interference of zero values ​​at night is eliminated, ensuring that the seasonal cycles, trends, and residual components obtained from subsequent decomposition have clear physical meanings, thereby improving the accuracy and interpretability of the predictions.

[0057] Step S302: Perform the first-level decomposition on the available power data to obtain the seasonal cycle component, trend component, and residual component.

[0058] This step employs a seasonal-trend decomposition method based on locally weighted regression to perform a first-level decomposition of the available daytime power sequence, obtaining seasonal periodic components, trend components, and residual components. In this embodiment, the preferred parameters for STL decomposition are: seasonal period length period = 144 (corresponding to 24 hours × 6 steps / hour, i.e., 10-minute resolution), seasonal smoother seasonal_smoother = 7, trend smoother trend_smoother = 1441, and robust mode (robust = True) enabled. This decomposition effectively separates signal components at different time scales, laying a solid foundation for subsequent extrapolation.

[0059] Step S303: Perform linear extrapolation on the trend component to obtain the predicted value of the trend component at multiple future time steps.

[0060] Step S304: Construct a daily cycle template based on the seasonal cycle components.

[0061] For each step_of_days∈{0,…,143}, a periodic template of length 144 is calculated using the daily periodic template function, as shown in equation (6): Equation (6) Step S305: Based on the daily cycle template, the seasonal cycle component is extrapolated using cycle reuse to obtain the predicted values ​​of the seasonal cycle component at multiple future time steps.

[0062] For the h-th step in the future, first calculate the position number of that step within the day. ∈[0,143], and then directly read the template value of the corresponding position from the daily cycle template as the prediction, as shown in equation (2): Equation (2) Where t represents the current time and h represents the sequence number of the future time step. This represents the predicted value of the seasonal cycle component at the h-th step in the future. This represents a daily periodic template function, whose input is the position number of the time step within a day. This indicates the position number of the h-th step in the future, meaning that the table can be looked up directly based on the position number of the future time step within a day.

[0063] This periodic extrapolation method utilizes the daily periodicity of photovoltaic power, which requires no training parameters and can accurately reproduce the power variation pattern of a typical day. It is especially suitable for stable patterns under sunny or cloudy weather conditions.

[0064] Step S306 involves performing a second-level decomposition on the residual components to obtain multiple intrinsic mode components. In some optional embodiments, the residual components contain high-frequency fluctuations not explained by trends and seasons, such as rapid changes in weather processes. To further separate signals from different frequency bands, this step applies variational mode decomposition (VMD) to the residual components. With the number of decompositions K=4, the penalty parameter α=2000, and the noise margin τ=0, solving the constrained variational problem yields four intrinsic mode components IMF_1 to IMF_4, corresponding to hourly weather processes, 10–30 minute cloud movement processes, minute-level fluctuations, and high-frequency noise, respectively.

[0065] Step S307: Extrapolate each intrinsic mode component using a first-order autoregressive model to obtain the predicted value of each intrinsic mode component at multiple future time steps.

[0066] In step S308, the predicted values ​​of the extrapolated seasonal cycle component, the predicted values ​​of the trend component, and the predicted values ​​of multiple intrinsic mode components are summed to obtain the prediction tensor of the full component reconstruction summation channel.

[0067] The calculation formula for the prediction tensor of the full component reconstruction summation channel is shown in Equation (5). Finally, the extrapolation results are combined into prediction tensors of 7 decomposed channels: seasonal channel (stl_seasonal), trend channel (stl_trend), four IMF channels (vmd_imf_1~vmd_imf_4) and full component reconstruction summation channel (decomp_sum, i.e., the sum of seasonal + trend + four IMFs).

[0068] Step S309: The predicted values ​​of the seasonal cycle component, the trend component, and each intrinsic mode component are used as the predicted values ​​of the seasonal channel, the trend channel, and each intrinsic mode component channel, respectively, thereby obtaining the predicted tensors of multiple decomposed channels.

[0069] In addition to the full-component reconstruction and summation channel, this invention also directly uses the extrapolation results of each independent component as separate prediction channels: a seasonal channel, a trend channel, and four IMF channels. This results in a total of seven decomposed channel prediction tensors (each tensor with a dimension representing the next 24 steps). These channels are then used as "virtual basis predictions" and input into the subsequent attention ensemble decision module, running alongside the deep model predictions. The attention mechanism dynamically learns the weights of each channel at different prediction steps. The advantage of using independent components as separate channels is that the ensemble model can flexibly choose which frequency band information to trust based on current weather conditions—for example, when clouds are moving rapidly, the weight of the high-frequency IMF channel may automatically increase; while during stable weather, the weight of the seasonal channel may dominate. This multi-channel decomposition-ensemble paradigm completely avoids the high cost of "training a deep model separately for each IMF" in traditional methods, while simultaneously improving the predictive adaptability.

[0070] This method achieves high-quality decomposition and analytical extrapolation of historical available power sequences. Daytime masking eliminates contamination from zero values ​​at night, ensuring the physical clarity of the decomposed components. The STL+VMD two-layer decomposition separates the signal into components with defined time scales, and analytical extrapolation requires no neural network training, resulting in extremely low computational cost. The final seven decomposed channel prediction tensors, together with the deep model predictions, constitute eleven basis predictors, providing a rich and complementary source of information for subsequent attention ensemble decision-making.

[0071] Figure 4 This is a schematic diagram illustrating the execution flow of the attention integration decision module in a photovoltaic power forecasting method based on satellite-ground collaborative observation according to an embodiment of the present invention. Figure 4 As shown, the process includes at least the following steps S401 to S412.

[0072] Step S401: Input the depth prediction sequences and prediction tensors into the attention fusion decision module. First, input the four depth prediction sequences obtained from the deep model layer (each sequence contains power prediction values ​​for the next 24 time steps) and the seven decomposition channel prediction tensors obtained from the decomposition extrapolation layer (each tensor also contains prediction values ​​for the next 24 steps) into the attention fusion decision module, thereby providing a rich information source for subsequent intelligent fusion.

[0073] Step S402: In the static prior branch of the attention integration decision module, obtain multiple deep prediction sequences.

[0074] Step S403: Obtain the static prior weights on the validation set through mean square error grid search.

[0075] Step S404: Perform weighted aggregation based on static prior weights to output a stable baseline.

[0076] Steps S402 to S404 constitute the static prior branch. In this branch, multiple depth prediction sequences are first acquired (step S402). Then, a fixed set of static prior weights is determined on the validation set using a mean squared error grid search (step S403). During the search, only the depth prediction sequences are used, and the weights of the decomposition channels are forced to zero to ensure that the initial ensemble results are not affected by unverified decomposition extrapolation errors. Finally, the depth prediction sequences are weighted and aggregated according to these weights to output a stable baseline (step S404). The significance of this baseline is that even if the dynamic attention branch is not yet well-learned, the ensemble performance is at least no worse than the static weighted ensemble of a pure deep model, thus ensuring a stable start to training.

[0077] Step S405: Combine the depth prediction sequence and the prediction tensor as the input tensor, and construct features based on the input tensor.

[0078] Step S406: The feature construction results are sequentially passed through a linear layer, layer normalization, modified linear unit, and dropout method to obtain the hidden representation.

[0079] Step S407: The attention weights are obtained by normalizing the logical values ​​of the linear layer output with the same dimension as the total number of base predictors and normalizing them using a normalized exponential function.

[0080] Step S408: The prediction values ​​of each base predictor at this time step are weighted and summed using attention weights to obtain the dynamically weighted prediction result.

[0081] Steps S405 to S408 constitute the dynamic attention branch. This branch utilizes all 11 base predictors (depth prediction sequence + prediction tensor). First, the input tensor, composed of the depth prediction sequence and the prediction tensor, is combined, and features are constructed based on this input tensor (step S405). Specifically, for each prediction time step h, the predicted values ​​of all base predictors at that step are extracted, and the statistical features of these predicted values ​​(mean, standard deviation, maximum, minimum, range) are calculated, while step-level position encoding is added. In addition to static prior mixed features, the feature vector is formed. .

[0082] Next, the feature construction results are sequentially processed through a linear layer, layer normalization, Corrected Linear Unit (ReLU), and dropout method to obtain a hidden representation (step S406). The hidden representation preferably has a dimension of 24, incorporating all relevant information for that time step. Then, a linear layer outputs logistic values ​​with the same dimension as the total number of base predictors, and these logistic values ​​are converted into attention weights using a normalized exponential function (step S407). Finally, the attention weights are used to weight and sum the predictions of each base predictor to obtain a dynamically weighted prediction result (step S408). The core advantage of this branch is that the attention weights are dynamically generated and can adaptively adjust according to sample features and time steps, thus flexibly allocating confidence levels under different weather conditions (sunny, cloudy, overcast) and different prediction distances (short-term, long-term), overcoming the fixed defects of static weights.

[0083] Step S409: Reuse the hidden representation of the dynamic attention branch.

[0084] Step S410: Generate residual gating coefficients using a linear layer and a preset activation function, respectively.

[0085] Step S411: Generate additive residuals through linear layers and modified linear units.

[0086] Steps S409 to S411 constitute the residual correction branch. This branch reuses the hidden representation already computed in the dynamic attention branch (step S409), avoiding redundant computation. Then, residual gating coefficients are generated through a linear layer and a preset activation function (Sigmoid) (step S410). These gating coefficients range from 0 to 1, and their linear layer bias is initialized to -1.0, resulting in an initial gs≈0.27. This means that the magnitude of the initial dynamic correction is suppressed to ensure training stability. Simultaneously, additive residuals are generated through another linear layer and a corrected linear unit (ReLU) (step S411). The additive residual is a learnable bias term used to fine-tune the final output.

[0087] Step S412: The stable baseline, dynamically weighted prediction results, residual gating coefficients, and additive residuals are calculated according to a preset formula to obtain the final power prediction result. One possible example of the preset formula is shown in equation (1): Equation (1) in, This represents the final power prediction result at the s-th prediction time step. Indicates a stable baseline. This indicates a dynamically weighted prediction. Represents the residual gating coefficient. This represents additive residuals.

[0088] The physical meaning of this formula is: based on static priors, dynamic attention prediction is allowed to deviate from the benchmark to a certain extent, with the deviation controlled by the residual gating coefficients; finally, a fine-tuning term is added. When the value is 0, it completely degenerates into a static weighted integration; when... When the gating value is 1, dynamic attention prediction is used entirely. As training progresses, the gating coefficient can adaptively increase or decrease, achieving an optimal balance between stability and flexibility. Additive residuals This further enhances the fitting capability.

[0089] Combining the calculation processes of the above branches, an optional example of its comprehensive output formula is shown in equation (7): Equation (7) in, Let M represent the final power prediction result at time step h, where M represents the number of deep heterogeneous basis models and C represents the number of prediction tensors. This represents the power prediction value of the m-th depth prediction sequence at the h-th time step. This represents the power prediction value of the c-th base predictor at time step h. The base predictor includes a depth prediction sequence and a prediction tensor. This represents the static prior weights of the m-th depth prediction sequence. This represents the dynamic attention weights of the c-th base predictor at the h-th time step. This represents the residual gating coefficient at the h-th time step. This represents the additive residual at the h-th time step.

[0090] pass Figure 4 The three-branch attention-integrated decision-making process shown can dynamically and adaptively fuse 11 heterogeneous base predictors. The static prior branch ensures that performance does not decline in the early stages of training; the dynamic attention branch enables flexible weight allocation at the sample and step levels, significantly improving the predictive adaptability under complex weather conditions; the residual correction branch prevents excessive deviation through a gating mechanism while providing fine-tuning terms. The three branches work together to achieve accurate forecasting of photovoltaic power generation.

[0091] Figure 5 This is an average attention weight heatmap of a photovoltaic power prediction method based on satellite-ground collaborative observation according to an embodiment of the present invention. Figure 5This diagram shows a heatmap of the attention weights of the attention ensemble decision module across all samples on the test set. The horizontal axis represents the prediction time step (step1, step3, ..., step23, i.e., from minute 10 to minute 230, showing a step size every 10 minutes), and the vertical axis represents the 11 base predictors: four deep models (BiLSTM-Attention, CNN-LSTM, Transformer, and TCN), and seven decomposition channels (decomp_sum, stl_seasonal, stl_trend, and vmd_imf_1 to vmd_imf_4). The color intensity of each cell indicates the average attention weight of that base predictor at the corresponding time step.

[0092] from Figure 5 The following patterns can be observed: (1) In the near-term prediction (step1 to step5, i.e., the next 10 to 50 minutes), the weights of the deep model (especially BiLSTM-Attention and Transformer) and the full component reconstruction summation channel (decomp_sum) are relatively high, indicating that the deep model and the overall reconstruction prediction have advantages in the ultra-short time. (2) As the prediction step size increases (step7 to step23), the weights of the seasonal channel (stl_seasonal) and the trend channel (stl_trend) gradually increase, reflecting that the contribution of daily periodicity and long-term trend in the medium to long term (about 1 to 4 hours) is more important. (3) Each intrinsic mode component (vmd_imf_1 to vmd_imf_4) also has a certain weight in some step sizes, indicating that high-frequency weather disturbances and cloud movement information are adaptively used to correct the prediction. This heatmap intuitively shows that the attention mechanism of this invention can dynamically adjust the contribution of each base predictor according to the prediction distance, making up for the shortcomings of traditional static weighted integration, and enhancing the interpretability of the model.

[0093] Figure 6 This is a scatter plot of the photovoltaic power forecast for the next 4 hours in the satellite-ground collaborative observation method of a photovoltaic power forecast according to an embodiment of the present invention. Figure 6 A scatter plot of aggregated predictions for the next 4 hours (24 time steps) on the test set is provided. The horizontal axis represents the measured available power, and the vertical axis represents the predicted available power, totaling 174,624 points (7276 samples × 24 steps in the test set). The dashed line in the plot is the 1:1 reference line, and the solid line is the linear regression fitted line y = 0.813x + 2.747. The evaluation metric is the coefficient of determination. =0.837, Root Mean Square Error (RMSE) = 7.338, Mean Absolute Error (MAE) = 4.390.

[0094] The scatter plot distribution shows that most points closely surround the 1:1 line, indicating that the method of this invention has high accuracy and low bias in predicting the overall power over the next 4 hours. These data fully demonstrate the improved prediction accuracy resulting from the synergy of satellite-ground collaborative observation, decomposition extrapolation, and attention-integrated decision-making in this invention.

[0095] Figure 7 This is a scatter plot of the photovoltaic power forecast for the next 30 minutes, generated by satellite-ground collaborative observation in a photovoltaic power forecasting method according to an embodiment of the present invention. Figure 7 This displays a scatter plot showing the single-step prediction of available power 30 minutes in the future, a key business focus, with 7276 test samples. The horizontal axis represents measured available power, and the vertical axis represents predicted available power. The evaluation metric is the coefficient of determination. =0.918, RMSE=5.086, MAE=2.841, regression line y=0.911x+1.515. The vast majority of points in the scatter plot closely follow the 1:1 line, indicating that the invention achieves extremely high accuracy in predicting the next 30 minutes, directly meeting the performance requirements of power dispatching for Automatic Generation Control (AGC) commands.

[0096] Figure 8 This is a schematic diagram of photovoltaic power forecasting using satellite-ground coordinated observation in another embodiment of the present invention. Figure 8 The figure shows a comparison curve between the power forecast and measured values ​​for a specific day (e.g., from 08:00 to 12:00). The solid line (or scatter plot) represents the measured available power, and the dashed line represents the rolling 4-hour forecast predicted by the present invention. The figure clearly shows a high degree of agreement between the predicted and measured curves in terms of trend and amplitude, especially during periods of rapid power increase and fluctuation. The predicted values ​​follow the measured changes promptly without significant lag or overshoot. This further verifies the robustness of the present invention under cloudy and abrupt weather conditions. Furthermore, the figure also indicates the confidence intervals of the predicted values ​​or the contributions of different decomposition channels, demonstrating the interpretability of the model.

[0097] In summary, Figures 5 to 8 From qualitative (heatmap weight distribution, time series comparison) and quantitative (scatter plot R) perspectives... 2 The results, based on two dimensions (RMSE and MAE), comprehensively demonstrate that the invention has achieved significant technical effects in terms of forecast accuracy, adaptive fusion capability, operational performance, and interpretability.

[0098] To more clearly illustrate the method of the present invention, a specific embodiment is provided for explanation: Implementation environment: Hardware: NVIDIA GeForce GTX 1060 (5GB VRAM) + Intel / AMD CPU + ≥8GB RAM.

[0099] Software: Python≥3.10, PyTorch≥2.0, NumPy, Pandas, scikit-learn, statsmodels, vmdpy.

[0100] Data: 2.5 years of historical observation data at 10-minute resolution for grid hub nodes including photovoltaic installations, approximately 140,000 records.

[0101] End-to-end steps: Step 1 (Data Acquisition): Collect raw data from the photovoltaic power station SCADA, satellite ground station, and station meteorological station, construct a 21-dimensional raw data column, and align the time sequence at a 10-minute resolution.

[0102] Step 2 (Feature Engineering): Complete the construction of 19-dimensional engineering-derived features (periodic encoding, hysteresis, differential, rolling, efficiency ratio), remove SAZ / SAA, and retain 40-dimensional input.

[0103] Step 3 (Data Partitioning): Divide the data into 80% / 10% / 10% training / validation / test sets according to time order. The normalizer is only fitted on the training set. Construct a sliding window sequence with lookback=72 steps and horizon=24 steps, and retain only the samples whose prediction window starts during the daytime period from 6:30 to 19:00.

[0104] Step 4 (Deep Model Training): Train four deep heterogeneous base models (BiLSTM-Attention, CNN-LSTM, Transformer, TCN) under a unified multi-regularization framework.

[0105] Step 5 (Signal Decomposition): Perform STL+VMD two-layer decomposition on the kygl sequence only during the training set period and only under the daytime mask to generate a daily periodic template (length 144), a trend sequence, and 4 IMF sequences.

[0106] Step 6 (Analytical Extrapolation): Generate 7 decomposed channel prediction tensors for each sample in the validation and test sets for the next 24 steps.

[0107] Step 7 (Stacker Training and Inference): Input the 11-channel tensor (4 depths + 7 decompositions) into AttentionStacker, train Stacker with composite loss (<5 min), and output the final prediction.

[0108] Step 8 (Post-processing and indicator calculation): Set the night forecast to zero, extract the forecast data for the next 30 minutes, and calculate the business assessment accuracy.

[0109] Through practical comparison, the method of this invention was fully validated on a dataset of 140,078 observations at a 10-minute resolution from a real power grid. Experimental results show that this invention significantly outperforms existing technologies in terms of accuracy, training cost, and overfitting suppression.

[0110] In summary, the photovoltaic power prediction method based on satellite-ground collaborative observation provided by this invention has the following technical advantages: 1. High-precision power prediction: By constructing multi-source observation features through satellite-ground collaboration, and integrating satellite remote sensing information on space clouds with ground-measured local power responses, a complementary "satellite-ground time" triangular information system is formed, significantly improving prediction accuracy. Experimental results show that the coefficient of determination for the aggregated prediction over the next 4 hours is [missing information]. The coefficient of determination reached 0.837, the root mean square error (RMSE) was 7.338, and the mean absolute error (MAE) was 4.390. For the forecast of the next 30 minutes, which is of most concern to power dispatch, the coefficient of determination was... With an accuracy of up to 0.918, a root mean square error (RMSE) of 5.086, a mean absolute error (MAE) of 2.841, and an operational performance accuracy of 83.54%, it can directly meet the high-precision requirements of automatic generator control (AGC) commands.

[0111] 2. Extremely low computational and training costs: This invention employs a two-layer STL+VMD signal decomposition under a daytime mask and directly extrapolates the prediction tensors of the decomposed channels through analytical methods such as periodic reuse, linear extrapolation, and first-order autoregressive models, eliminating the need to train a separate deep neural network for each intrinsic mode component. The overall training workload is independent of the number of intrinsic mode components, requiring only the training of four deep heterogeneous base models and a lightweight attention stacker at once. On a conventional graphics processor (such as GTX1060), the total training time can be controlled within 35 minutes, and the single-sample inference time is less than 50 milliseconds, meeting the requirements of real-time rolling forecasting in engineering.

[0112] 3. Strong robustness and physical interpretability: Daytime masking eliminates approximately 50.9% of zero values ​​from nighttime, preventing spurious oscillations from contaminating the decomposition results. This ensures that the seasonal cycle components, trend components, and intrinsic model components obtained from the decomposition have clear physical meanings (e.g., hourly weather processes, 10-30 minute cloud movement, minute-level fluctuations, and high-frequency noise). The attention weight heatmap output by the attention integration decision module visually demonstrates the contribution of each base predictor at different prediction steps, enhancing the model's interpretability and facilitating understanding of the prediction basis by operations and maintenance personnel.

[0113] 4. Excellent maintainability and portability: The deep model layer and the decomposition-integration layer are logically decoupled in the training pipeline. Adjustments to the integration layer (Stacker) or decomposition parameters do not require retraining the deep model, reducing the model iteration cycle from "days" to "hours". Furthermore, this method is not dependent on specific satellite models; any geostationary meteorological satellite that can provide albedo and brightness temperature channels (such as Sunflower and Fengyun-4) can be used as the satellite-end input. The ground end only requires industry-standard radiometers and SCADA data interfaces, facilitating rapid migration and deployment across different photovoltaic sites.

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

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

[0116] This embodiment also provides a computer program product 10, a computer-readable storage medium 20, and a computer device 30. Figure 9 This is a schematic diagram of a computer program product 10 according to an embodiment of the present invention. Figure 10 This is a schematic diagram of a computer-readable storage medium 20 according to an embodiment of the present invention. Figure 11 This is a schematic diagram of a computer device 30 according to an embodiment of the present invention. The computer program product 10 includes a computer program 11, which, when executed by the processor 32, implements the steps of the above-described satellite-ground collaborative observation photovoltaic power forecasting method. A computer-readable storage medium 20 stores the above-described computer program 11, which, when executed by the processor 32, implements the steps of the satellite-ground collaborative observation photovoltaic power forecasting method of any of the above-described embodiments. The computer device 30 may include a memory 31, a processor 32, and the computer program 11 stored in the memory 31 and running on the processor 32.

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

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

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

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

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

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

Claims

1. A method for predicting the power of photovoltaic power plants using satellite-ground collaborative observation, comprising: Acquire satellite-based and ground-based observation data of the location of the photovoltaic power station; Based on the ground-based observation data, engineering-derived features are constructed, and the satellite-based observation data, the ground-based observation data, and the engineering-derived features are fused into multi-source observation features. Multiple pre-trained deep heterogeneous basis models are used to process the multi-source observation features to obtain multiple deep prediction sequences. The deep prediction sequences contain power prediction values ​​for multiple future time steps. The deep heterogeneous basis models are used to extract temporal dependencies from the multi-source observation features and map them to obtain power prediction values ​​for multiple future time steps. Daytime masking and two-layer signal decomposition are performed on the historical available power sequence, and the prediction tensors of multiple decomposition channels are obtained by analytical extrapolation. The depth prediction sequence and the prediction tensor are input into the attention fusion decision module, which outputs the final power prediction result. The attention fusion decision module is used to fuse the depth prediction sequence and the prediction tensor through an attention mechanism to generate the final power prediction result.

2. The photovoltaic power forecasting method based on satellite-ground collaborative observation according to claim 1, wherein, The steps of performing daytime masking and two-layer signal decomposition on the historical available power sequence, and obtaining the prediction tensor of the decomposed channel through analytical extrapolation, include: Construct daytime masking rules, and filter the available power sequence in the historical available power sequence according to the daytime masking rules to obtain the available power sequence in the preset daytime period; The available power sequence is decomposed into a first-level decomposition to obtain seasonal periodic components, trend components, and residual components. The first-level decomposition includes a seasonal trend decomposition method. Linear extrapolation is performed on the trend component to obtain the predicted value of the trend component at multiple future time steps; A daily cycle template is constructed based on the seasonal cycle components; Based on the daily cycle template, the seasonal cycle component is extrapolated using cycle reuse to obtain the predicted value of the seasonal cycle component in the future multiple time steps. The residual components are subjected to a second-level decomposition to obtain multiple intrinsic mode components. The second-level decomposition includes variational mode decomposition. Each intrinsic mode component is extrapolated using a first-order autoregressive model to obtain the predicted value of each intrinsic mode component at multiple future time steps. The predicted values ​​of the extrapolated seasonal cycle component, the predicted value of the trend component, and the predicted values ​​of multiple intrinsic mode components are summed to obtain the prediction tensor of the full component reconstruction summation channel. The predicted values ​​of the seasonal cycle component, the trend component, and each of the intrinsic mode components are used as the predicted values ​​of the seasonal channel, the trend channel, and each intrinsic mode component channel, respectively, thereby obtaining the predicted tensors of multiple decomposed channels.

3. The photovoltaic power forecasting method based on satellite-ground collaborative observation according to claim 2, wherein, The intrinsic model components obtained from the variational model decomposition include hourly weather processes, 10- to 30-minute cloud movement processes, minute-level fluctuations, and high-frequency noise. The decomposition channels include a seasonal channel, a trend channel, channels for each of the intrinsic mode components, and a full component reconstruction summation channel. The seasonal channel corresponds to the predicted value of the seasonal cycle component at multiple future time steps, the trend channel corresponds to the predicted value of the trend component at multiple future time steps, and the predicted value of the full component reconstruction summation channel is obtained by summing the predicted values ​​of the seasonal cycle component, the trend component, and each of the intrinsic mode components.

4. The photovoltaic power forecasting method based on satellite-ground coordinated observation according to claim 2, wherein, The calculation formula for periodic extrapolation of the seasonal periodic components is as follows: Where t represents the current time and h represents the sequence number of the future time step. This represents the predicted value of the seasonal cycle component at the h-th step in the future. This represents a daily periodic template function, whose input is the position number of the time step within a day. Indicates the position number at the h-th step in the future; The formula for linear extrapolating the trend components is as follows: in, This represents the predicted value of the trend component at the h-th time step in the future, where a and b are the slope and intercept of the linear extrapolation, respectively, and are obtained by fitting the trend component values ​​of the last multiple time steps using the least squares method. The formula for extrapolating each of the intrinsic mode components using a first-order autoregressive model is as follows: in, This represents the predicted value of the k-th intrinsic mode component at the h-th step in the future. The first-order autoregressive coefficient of the k-th intrinsic mode component is estimated from the samples of the last few time steps of this component. Let K be the average of the samples of the k-th intrinsic mode component over the last multiple time steps, where K is the number of intrinsic mode components. The formula for calculating the prediction tensor of the full component reconstruction summation channel is as follows: in, This represents the prediction tensor of the full component reconstruction summation channel.

5. The photovoltaic power forecasting method based on satellite-ground coordinated observation according to claim 1, wherein, The step of inputting the depth prediction sequence and the prediction tensor into the attention integration decision module, and having the attention integration decision module output the final power prediction result, includes: The depth prediction sequence and the prediction tensor are combined as the input tensor of the attention integration decision module; In the static prior branch of the attention integration decision module, multiple deep prediction sequences are obtained, and static prior weights are obtained on the validation set through mean square error grid search. The static prior weights are then weighted and aggregated to output a stable baseline. In the dynamic attention branch of the attention integration decision module, features are constructed based on the input tensor, and the feature construction results are sequentially passed through a linear layer, layer normalization, modified linear unit, and dropout method to obtain a hidden representation. The linear layer outputs a logical value with the same dimension as the total number of base predictors, and the attention weights are obtained by normalizing through a normalized exponential function. The attention weights are then used to perform a weighted summation of the prediction values ​​of each base predictor at this time step to obtain a dynamic weighted prediction result. In the residual correction branch of the attention integration decision module, the hidden representation of the dynamic attention branch is reused, and residual gating coefficients are generated through a linear layer and a preset activation function, respectively, and additive residuals are generated through a linear layer and a modified linear unit. The stable baseline, the dynamically weighted prediction result, the residual gating coefficient, and the additive residual are calculated according to a preset formula to obtain the final power prediction result.

6. The photovoltaic power forecasting method based on satellite-ground collaborative observation according to claim 5, wherein, The formula for calculating the final power prediction result is as follows: in, This represents the final power prediction result at time step h, where M represents the number of deep heterogeneous basis models and C represents the number of prediction tensors. This represents the power prediction value of the m-th depth prediction sequence at the h-th time step. This represents the power prediction value of the c-th base predictor at time step h. The base predictor includes a depth prediction sequence and a prediction tensor. This represents the static prior weights of the m-th depth prediction sequence. This represents the dynamic attention weights of the c-th base predictor at the h-th time step. This represents the residual gating coefficient at the h-th time step. This represents the additive residual at the h-th time step.

7. The photovoltaic power forecasting method based on satellite-ground coordinated observation according to claim 1, wherein, The satellite-based observation data includes albedo data and brightness temperature data for multiple bands from geostationary meteorological satellites. The ground-based observation data includes total radiation data measured by the station's radiometer and multi-channel measured power data from the site. The engineering-derived features include the following features: One or more of the following: periodic coding features, aggregate power features, multi-scale lag features, ultra-short-term rate of change features, rolling statistical features, and efficiency ratio features.

8. The photovoltaic power forecasting method based on satellite-ground collaborative observation according to claim 7, wherein, The satellite observation data includes albedo data in 6 visible or near-infrared bands and brightness temperature data in 10 infrared bands, totaling 16 dimensions of data. The ground-based observation data includes 1-dimensional total radiation data and 4 channels of measured power data, for a total of 5 dimensions. The engineering-derived features specifically include: 6-dimensional periodic coding features, 1-dimensional aggregated power features, 5-dimensional multi-scale lag features, 3-dimensional ultra-short-term change rate features, 2-dimensional rolling statistical features, and 1-dimensional efficiency ratio features, for a total of 19 features.

9. A computer program product comprising a computer program that, when executed by a processor, implements the steps of the photovoltaic power forecasting method based on satellite-ground collaborative observation as described in any one of claims 1 to 8.

10. A computer device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the computer program to implement the steps of the photovoltaic power forecasting method for satellite-ground collaborative observation according to any one of claims 1 to 8.

Images