Photovoltaic power rolling prediction method and system fusing weather warning and cloud cluster remote sensing

Abstract

This invention discloses a method and system for rolling photovoltaic power prediction that integrates meteorological early warning and cloud remote sensing, belonging to the field of new energy power prediction technology. This invention acquires multi-source data at both long and short scales in a layered manner, simultaneously obtaining regular operational data and high-frequency data on short-term irradiance mutations. It constructs a weighted model that fuses meteorological early warning levels and cloud motion vectors, and combines this with a dynamic correction model input feature vector based on irradiance mutation gradients to achieve adaptive adjustment of features across multiple meteorological scenarios. An Informer encoder-multi-scale LSTM fusion model is built to consider both long-term power trends and short-term mutation fluctuations, outputting high-precision initial predicted power. This invention solves the technical pain points of traditional photovoltaic prediction technologies, such as delayed prediction of meteorological mutations and easy model drift, significantly improving the prediction accuracy for complex scenarios such as rainfall, thunderstorms, and rapid cloud obstruction. It exhibits strong long-term operational stability and can be widely adapted to various ultra-short-term power prediction scenarios for photovoltaic power plants and real-time grid dispatching.

Description

Technical Field

[0001] This invention belongs to the field of new energy power prediction technology, specifically involving a method and system for rolling prediction of photovoltaic power that integrates meteorological early warning and cloud remote sensing. Background Technology

[0002] Ultra-short-term photovoltaic (PV) power forecasting is a core foundational technology for ensuring the stable operation of new power systems, enhancing the absorption capacity of new energy sources, and supporting real-time grid dispatch. PV output is highly random, volatile, and prone to sudden changes, making it extremely susceptible to weather factors such as cloud cover, sudden weather changes, strong winds, and rainfall. Accurate ultra-short-term power forecasting is crucial for avoiding power supply-demand imbalances in the grid, reducing curtailment of solar power, and optimizing power dispatch strategies.

[0003] Existing ultra-short-term photovoltaic power prediction technologies have significant technical shortcomings: First, most existing prediction models rely solely on historical power data of power plants and conventional meteorological data for modeling, without incorporating official meteorological warning levels and satellite cloud imagery remote sensing data. They cannot predict sudden meteorological events such as thunderstorms, heavy rain, severe convection, and rapid cloud obstruction in advance, and can only passively adapt to data changes after photovoltaic output fluctuates. The prediction lag is significant, and the prediction accuracy is greatly reduced in scenarios of sudden meteorological changes.

[0004] Secondly, existing technologies lack differentiated feature weight control mechanisms, making it impossible to distinguish the different impacts of different levels of weather warnings on photovoltaic power output. For multiple weather scenarios such as no warning, blue, yellow, orange, and red, a uniform feature weight processing method is used, which results in over-reliance on meteorological data in non-warning scenarios and insufficient meteorological feature weights in extreme warning scenarios, leading to poor scenario adaptability.

[0005] Third, traditional prediction models mostly adopt a fixed-period update mode, with a low update frequency, which cannot match the real-time dispatch requirements of the power grid (15-30 minutes) and lacks a model anti-drift optimization mechanism. During the long-term operation of photovoltaic power plants, component aging and environmental climate change will cause continuous drift in the distribution of operating data. Traditional fixed-parameter models cannot adaptively iterate and update, resulting in a continuous decline in prediction accuracy and insufficient stability over long-term operation.

[0006] Fourth, existing time series prediction models have limited time series feature extraction capabilities and cannot take into account both short-term irradiance abrupt changes and long-term power change trends. They also have poor adaptability to rapid irradiance fluctuations at the 5-minute level and slow power changes at the 30-minute level, making it difficult to accurately capture the multi-scale fluctuation characteristics of photovoltaic power under complex weather conditions. Summary of the Invention

[0007] To address the problems mentioned in the background art, this invention provides a method and system for rolling photovoltaic power prediction that integrates meteorological early warning and cloud remote sensing, in order to solve the problems of delayed forecasting of sudden weather events, poor scene adaptability, easy model drift, weak multi-scale fluctuation capture capability, and mismatch between update mechanism and power grid dispatch in the prior art.

[0008] To achieve the above objectives, the present invention provides the following technical solution:

[0009] A method for rolling forecasting of photovoltaic power that integrates weather early warning and cloud remote sensing includes the following steps:

[0010] S1: Multi-source heterogeneous data hierarchical acquisition: Obtain conventional power plant operation data and meteorological data at the first preset time granularity, including historical power data, real-time temperature, irradiance data, cloud movement vectors obtained from satellite cloud image inversion and meteorological warning levels, and simultaneously collect irradiance abrupt change samples at the second preset time granularity to construct a long and short dual-scale dataset; the second preset time granularity is smaller than the first preset time granularity.

[0011] S2: Dynamic correction of meteorological-power characteristics: Meteorological feature vector and power feature vector are constructed based on the data obtained in S1; the warning impact weight of the fused cloud vector is calculated based on the meteorological warning level and cloud movement vector obtained in S1, and the meteorological feature vector and power feature vector are weighted and corrected by combining the gradient of sudden change in irradiance, so as to obtain the corrected model input feature vector.

[0012] S3: Multi-scale fusion model power prediction: The corrected model input feature vector obtained in S2 is input into the Informer encoder-multi-scale LSTM fusion deep learning model. The output results of the two models are weighted by the optimal fusion weight to obtain the initial predicted power of photovoltaic power.

[0013] S4: Edge-Cloud Collaborative Rolling Incremental Optimization: Access measured power plant operation data and real-time meteorological data according to a preset rolling cycle, and calculate the prediction mean square error by combining the initial photovoltaic power prediction obtained in S3; when the mean square error is greater than the preset error threshold, start the incremental training mechanism, and dynamically update the model parameters through the adaptive learning rate formula to complete the model anti-drift rolling optimization.

[0014] A photovoltaic power rolling forecasting system integrating meteorological early warning and cloud remote sensing includes a data acquisition module, a satellite cloud image analysis module, a feature correction module, a fusion forecasting module, and an edge-cloud collaborative rolling update module;

[0015] The data acquisition module's output is connected to the satellite cloud image analysis module's input and the feature correction module's input, respectively. The satellite cloud image analysis module's output is connected to the feature correction module's input, the feature correction module's output is connected to the fusion prediction module's input, the fusion prediction module's output is connected to the edge-cloud collaborative rolling update module's input, and the edge-cloud collaborative rolling update module's parameter output is connected to the fusion prediction module's parameter update.

[0016] The data acquisition module is used to collect conventional power plant operation data at a first preset time granularity and radiation intensity mutation data at a second preset time granularity, and simultaneously acquire meteorological early warning data and satellite cloud image remote sensing data.

[0017] The satellite cloud image analysis module is used to invert cloud cluster motion vectors and cloud coverage parameters based on satellite cloud image data, and to quantify cloud cluster movement and occlusion characteristics.

[0018] The feature correction module is used to calculate the impact weight of meteorological warnings and the gradient of sudden changes in irradiance, and to dynamically correct the meteorological and power fusion feature vector.

[0019] The fusion prediction module is used to carry an Informer encoder-multiscale LSTM fusion model and output the initial predicted power of photovoltaic power with weights.

[0020] The edge-cloud collaborative rolling update module is used to monitor prediction errors in real time, and after triggering a preset error threshold, it performs incremental training and adaptive learning rate updates to achieve model anti-drift rolling optimization.

[0021] Compared with the prior art, the beneficial effects of the present invention are:

[0022] 1. This invention uses a weighting mechanism that integrates meteorological warning levels and cloud movement vectors to quantify the impact of multi-level meteorological warnings and cloud shading on photovoltaic power output. It replaces the traditional unified weighting processing mode, strengthens the dominant role of historical power characteristics in scenarios without warnings, strengthens the weight of meteorological remote sensing features in high-level extreme warning scenarios, accurately adapts to all meteorological scenarios, significantly reduces prediction deviations under sudden weather conditions, can predict the trend of sudden drop in irradiance, and eliminates prediction lag problems.

[0023] 2. This invention designs a dual-scale data acquisition mechanism that simultaneously captures power change trends and irradiance mutation details. It is combined with a multi-scale LSTM model to adapt to power fluctuation characteristics at different time scales, solving the technical defects of traditional models that cannot take into account both fast and slow changing conditions, and significantly improving the prediction accuracy in complex weather and weather switching scenarios.

[0024] 3. This invention constructs a rolling update + incremental anti-drift optimization mechanism, which matches the real-time scheduling update frequency of the power grid. By adaptively triggering model parameter iteration through error threshold, it effectively offsets the accuracy decay problem caused by photovoltaic module aging and environmental drift, ensures the long-term operational stability of the model, and solves the pain points of traditional models such as fixed parameters, delayed updates, and easy accuracy drift.

[0025] 4. This invention adopts the Informer-multiscale LSTM dual-model fusion architecture, adds a dedicated attention head for irradiance mutations, enhances the ability to extract short-term mutation features, and the weighted fusion of the two models takes into account both long-term trend prediction and short-term mutation capture. The prediction stability and robustness are significantly better than a single prediction model.

[0026] 5. This invention adopts an edge-cloud collaborative deployment architecture, which can compress the overall prediction latency of the system. The entire process of data acquisition, feature correction, model prediction, and parameter update is automated without manual intervention, and it is suitable for ultra-short-term power prediction and grid dispatching application scenarios of various distributed and centralized photovoltaic power plants. Attached Figure Description

[0027] Figure 1 This is a schematic diagram illustrating the specific process of this application. Detailed Implementation

[0028] To facilitate understanding of the technical content of this invention by those skilled in the art, the invention will be further described in detail below with reference to the accompanying drawings and specific examples. It should be understood that the specific examples described herein are merely illustrative and not intended to limit the scope of the invention.

[0029] A method for rolling forecasting of photovoltaic power that integrates weather early warning and cloud remote sensing, such as Figure 1 As shown, it includes the following steps:

[0030] S1: Multi-source heterogeneous data hierarchical acquisition: acquiring historical power data P hist Real-time temperature T, irradiance G, cloud motion vector V retrieved from satellite cloud imagery, and meteorological warning level W are used to construct a dual-scale dataset, with a conventional prediction data time granularity of 15 minutes and simultaneous high-frequency collection of 1-minute irradiance abrupt change samples.

[0031] S2: Dynamic correction of meteorological and power characteristics: The warning impact weight of the fused cloud vector is calculated by a preset weight formula, and the meteorological and power fused feature vector is corrected by combining the short-term irradiance change gradient, so as to realize the adaptive allocation of feature weights in sudden meteorological scenarios.

[0032] The formula for calculating the impact weight of early warning is:

[0033] ;

[0034] The formula for calculating eigenvector correction is:

[0035] ;

[0036] in, The weighting for the impact of early warning is set to an upper limit of 1.0; W represents the weather warning level; |V| represents the magnitude of the cloud cluster motion vector. Input feature vectors into the corrected model; This is a meteorological feature vector; G is the power eigenvector; G' is the gradient of sudden changes in irradiance.

[0037] The meteorological warning level W is defined as follows: no warning W=0, blue warning W=1, yellow warning W=2, orange warning W=3, and red warning W=4; the cloud movement vector V is in km / h.

[0038] Meteorological feature vector Includes real-time temperature T, real-time irradiance G, weather warning level W, cloud movement vector V, cloud cover C; power feature vector Includes historical power P hist Historical power change rate dP hist / dt, the power change rate calculation window is 30 minutes; the irradiance abrupt change gradient G'=dG / dt, the gradient calculation window is 5 minutes.

[0039] S3: Power prediction using a multi-scale fusion model: Construct an Informer encoder-multi-scale LSTM fusion deep learning model, and obtain the initial predicted power of photovoltaic power by weighting the dual model output results with optimal fusion weights.

[0040] Before performing the initial predicted power fusion calculation, the feature vector Xmodified, which has been corrected by the meteorological warning weights, will be used as a unified input feature and fed in parallel into the two prediction branches of the Informer encoder network and the multi-scale LSTM network:

[0041] 1. The modified feature vector Xmodified is fully input into the Informer encoder. After the encoder performs temporal attention encoding and long-period temporal dependency mining operation, the long-term global photovoltaic power prediction value PInformer is output.

[0042] The same set of modified feature vectors Xmodified are synchronously input into a multi-scale LSTM network. The network divides the input features into multiple time-granular subsequences to extract local short-term fluctuation features. After fusing the time-series features at each scale, it outputs the photovoltaic power prediction value PLSTM under short-term abrupt change conditions.

[0043] Two sets of independent prediction results obtained based on the same modified feature vector are then fused into the following initial prediction power fusion formula to complete the weighted fusion operation:

[0044] ;

[0045] Wherein, λ is the model fusion weight, with a value ranging from 0.3 to 0.7. The optimal value is determined through 7-fold time series hierarchical cross-validation, and the validation set accounts for 20% of the total data. Output the prediction results for the Informer encoder; The prediction results are output by the multi-scale LSTM.

[0046] The Informer encoder and multi-scale LSTM model parameters are configured as follows: the Informer encoder has 3 layers, 8 attention heads, and a hidden layer dimension of 256, with a separate attention head added for short-term irradiance mutations; the multi-scale LSTM model has 3 layers, with hidden layer dimensions configured as 256, 128, and 64 respectively, corresponding to 5-minute, 15-minute, and 30-minute time series scales, and the overall dropout rate of the model is set to 0.2.

[0047] S4: Edge-cloud collaborative rolling incremental optimization: Every 15 minutes, the latest measured operation data and meteorological data are accessed to calculate the prediction mean square error (MSE); when the MSE > 3.5% error threshold, the incremental training mechanism is started, and the model parameters are dynamically updated through the adaptive learning rate formula to complete the model anti-drift rolling optimization.

[0048] The formula for calculating the mean square error is:

[0049] ;

[0050] The adaptive incremental learning rate update formula is:

[0051] ;

[0052] Where N is the number of prediction samples; Let be the actual photovoltaic power of the i-th sample; Let η be the photovoltaic predicted power of the i-th sample; η(t) be the dynamic learning rate at time t; η0 be the initial learning rate, ranging from 0.001 to 0.01; μ be the single-iteration decay coefficient, ranging from 0.05 to 0.11 / iteration; Δη be the incremental training compensation term, used to offset the data drift error during long-term model operation; the maximum number of rolling updates is 16, and the number of single incremental training steps is ≤10.

[0053] The number of predicted samples N ranges from 32 to 64. The data preprocessing process includes outlier removal and normalization. The Min-Max normalization formula is used to normalize the regular features to the [0,1] interval. Separate weighted normalization is performed on the 1-minute irradiance mutation samples to strengthen the weight of short-term mutation features.

[0054] The multi-source data acquisition methods are as follows: meteorological data is acquired in real time through the meteorological bureau's API interface, with a data update delay of ≤5 minutes; satellite cloud image data is collected every 10 minutes and the cloud movement vector parameters are obtained through inversion; historical power and real-time operating power data are retrieved from the photovoltaic power station's SCADA system, and the data integrity of the entire station is ≥99%.

[0055] A photovoltaic power rolling forecasting system integrating meteorological early warning and cloud remote sensing includes a data acquisition module, a satellite cloud image analysis module, a feature correction module, a fusion forecasting module, and an edge-cloud collaborative rolling update module;

[0056] The data acquisition module's output is connected to the satellite cloud image analysis module's input and the feature correction module's input, respectively. The satellite cloud image analysis module's output is connected to the feature correction module's input, the feature correction module's output is connected to the fusion prediction module's input, the fusion prediction module's output is connected to the edge-cloud collaborative rolling update module's input, and the edge-cloud collaborative rolling update module's parameter output is connected to the fusion prediction module's parameter update.

[0057] The data acquisition module is used to collect conventional power plant operation data at a first preset time granularity and radiation intensity mutation data at a second preset time granularity, and simultaneously acquire meteorological early warning data and satellite cloud image remote sensing data.

[0058] The satellite cloud image analysis module is used to invert cloud cluster motion vectors and cloud coverage parameters based on satellite cloud image data, and to quantify cloud cluster movement and occlusion characteristics.

[0059] The feature correction module is used to calculate the impact weight of meteorological warnings and the gradient of sudden changes in irradiance, and to dynamically correct the meteorological and power fusion feature vector.

[0060] The fusion prediction module is used to carry an Informer encoder-multiscale LSTM fusion model and output the initial predicted power of photovoltaic power with weights.

[0061] The edge-cloud collaborative rolling update module is used to monitor prediction errors in real time, and after triggering a preset error threshold, it performs incremental training and adaptive learning rate updates to achieve model anti-drift rolling optimization.

[0062] The invention will now be described with reference to specific examples.

[0063] This invention selects a 10MW distributed photovoltaic power station as the verification object. The photovoltaic modules are monocrystalline silicon photovoltaic modules, and the inverter model is Huawei SUN2000-60KTL. An edge data acquisition device and a cloud model training platform are built. The deep learning framework used is PyTorch 1.12, and the programming language is Python 3.9. The power station's annual full-condition operation data are selected, covering all typical scenarios such as sunny days without warning, cloudy days, yellow rainstorm warnings, orange gale warnings, and red rainstorm extreme warnings. Multi-scenario comparative verification is carried out to effectively verify the feasibility and superiority of the technical solution of this invention.

[0064] Example 1: No warning, sunny day conditions;

[0065] Under this operating condition, there is no weather warning, W=0, the weather is clear with no cloud cover, the cloud movement vector V=0km / h, and the cloud cover C=5%. The real-time temperature is T=28℃, the real-time irradiance is G=800W / m², the historical stable power is 9.5MW, and the 5-minute irradiance gradient is G'=0.2W / (m²·min).

[0066] Calculate the impact weight ω of the early warning W =0.1 + 0.2 × 0 + 0.15 × 0 = 0.1, low weights reinforce the dominant role of historical power characteristics. Construct meteorological feature vector X. weather =[28,800,0,0,5]、Power feature vector X power =[9.5,0.2], substituting into the correction formula yields the optimized fused features. The optimal fusion weight λ=0.4 is determined through 7-fold cross-validation. The Informer model outputs a predicted power of 9.4kW, the multi-scale LSTM model outputs a predicted power of 9.6kW, and the weighted fusion yields an initial predicted power of 9.52kW.

[0067] Rolling updates are performed every 15 minutes. The measured power in this round was 9.48kW. With a sample size of N=32, the calculated MSE ≈ 0.0005% < 3.5% error threshold, so incremental training and parameter updates are not triggered. After 16 consecutive rolling iterations, the prediction error stabilized at 2%-3%, demonstrating excellent prediction accuracy and good stability under clear weather steady-state conditions.

[0068] Example 2: Yellow alert for moderate rainfall;

[0069] A yellow weather warning is issued for this operating condition. W=2, moderate rainfall accompanied by rapidly moving clouds, cloud movement vector V=6km / h, cloud cover C=85%. Real-time temperature T=22℃, real-time irradiance G=300W / m², historical power 5.2MW, 5-minute irradiance abrupt change gradient G'=-3.2W / (m²·min).

[0070] Calculate the theoretical early warning weight ω W =1.4, taking the upper limit of 1.0, fully enhances the weight of meteorological remote sensing features to adapt to rainfall shading conditions. The corrected feature vector is mainly based on meteorological abrupt change features, weakening the influence of historical power data. The optimal fusion weight λ=0.5, the Informer model output is 4.8kW, the multi-scale LSTM model output is 5.0kW, and the fused initial predicted power is 4.9kW.

[0071] The measured power in this round was 4.2kW, with a sample size of N=36. The calculated MSE was approximately 3.6% > 3.5%, triggering incremental rolling optimization. The initial learning rate η0=0.005, attenuation coefficient μ=0.08, and incremental compensation term Δη=0.0005 were set. Iterative calculation yielded a real-time learning rate η=0.0048, completing the model parameter update. After the update, the predicted power for the next cycle was 4.3kW, with a deviation from the measured power ≤2%, effectively correcting the prediction error caused by sudden weather changes.

[0072] Example 3: Red alert for extreme rainstorm conditions;

[0073] This operating condition triggers a red extreme weather warning: W=4, indicating heavy rain and severe convective weather. The entire field area is covered by clouds, with a cloud cover rate of C=100% and a cloud movement vector V=12km / h. The real-time temperature is T=20℃, the real-time irradiance is G=50W / m², the historical power is 1.5MW, and the 5-minute irradiance abrupt change gradient is G'=-8.5W / (m²·min).

[0074] Calculate the theoretical early warning weight ω W =2.7, taking the upper limit of 1.0, the model fully responds to the characteristics of extreme weather changes. The optimal fusion weight λ=0.6, the Informer model output is 1.2kW, the multi-scale LSTM model output is 1.0kW, and the fused initial prediction power is 1.12kW.

[0075] The measured power in this round was 0.8kW, with a sample size of N=40. The MSE ≈ 3.8% > 3.5% error threshold, so incremental training optimization was immediately initiated. The initial learning rate η0 = 0.008, attenuation coefficient μ = 0.06, and incremental compensation term Δη = 0.001 were set. After the update, the real-time learning rate η = 0.0085. The secondary predicted output power was 0.85kW, with a deviation of only 0.05kW from the measured power. This demonstrates high-precision prediction even under extremely harsh operating conditions, fully meeting the stringent requirements of power grid dispatch.

[0076] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for rolling forecasting of photovoltaic power integrating meteorological early warning and cloud cluster remote sensing, characterized in that, Includes the following steps: S1: Multi-source heterogeneous data hierarchical acquisition: Obtain conventional power plant operation data and meteorological data at the first preset time granularity, including historical power data, real-time temperature, irradiance data, cloud movement vectors obtained from satellite cloud image inversion and meteorological warning levels, and simultaneously collect irradiance abrupt change samples at the second preset time granularity to construct a long and short dual-scale dataset; the second preset time granularity is smaller than the first preset time granularity. S2: Dynamic correction of meteorological-power characteristics: Meteorological feature vector and power feature vector are constructed based on the data obtained in S1; the warning impact weight of the fused cloud vector is calculated based on the meteorological warning level and cloud movement vector obtained in S1, and the meteorological feature vector and power feature vector are weighted and corrected by combining the gradient of sudden change in irradiance, so as to obtain the corrected model input feature vector. S3: Multi-scale fusion model power prediction: The corrected model input feature vector obtained in S2 is input into the Informer encoder-multi-scale LSTM fusion deep learning model. The output results of the two models are weighted by the optimal fusion weight to obtain the initial predicted power of photovoltaic power. S4: Edge-Cloud Collaborative Rolling Incremental Optimization: Access measured power plant operation data and real-time meteorological data according to a preset rolling cycle, and calculate the prediction mean square error by combining the initial photovoltaic power prediction obtained in S3; when the mean square error is greater than the preset error threshold, start the incremental training mechanism, and dynamically update the model parameters through the adaptive learning rate formula to complete the model anti-drift rolling optimization.

2. The photovoltaic power rolling prediction method integrating meteorological early warning and cloud cluster remote sensing according to claim 1, characterized in that, In S2, the formula for calculating the impact weight of the early warning is: ； in, To mitigate the impact of early warnings, the upper limit is set to 1.

0. This is a quantitative value for the meteorological warning level. Meteorological warnings are divided into multiple levels with corresponding incremental quantitative values. The magnitude of the cloud's motion vector; The formula for calculating eigenvector correction is: ； in, Input feature vectors to the corrected model. The meteorological feature vector includes real-time temperature, real-time irradiance, meteorological warning level, cloud movement vector, and cloud cover rate. This is a power feature vector, which includes historical power and the historical power change rate calculated based on a preset power change window. The gradient of irradiance abrupt change is calculated based on a preset gradient calculation window.

3. The photovoltaic power rolling prediction method integrating meteorological early warning and cloud cluster remote sensing according to claim 1, characterized in that, In S3: The initial predicted power fusion formula is: ； in The model fusion weights are set to values ​​within a preset range. The optimal values ​​are determined by time series stratified cross-validation with a preset number of folds. The validation set accounts for a preset proportion of the total data. Output the prediction results for the Informer encoder; Output prediction results for multi-scale LSTM; The Informer encoder is configured with a multi-layer structure, with multiple attention heads and a separate attention head for short-term radiation intensity abrupt changes. The hidden layer is set to a preset dimension. The multi-scale LSTM model is configured with a multi-layer structure, with the hidden layer dimension decreasing layer by layer, corresponding to multiple time scales of different durations. The model as a whole is set with a preset dropout rate.

4. The photovoltaic power rolling forecasting method integrating meteorological early warning and cloud cluster remote sensing according to claim 1, characterized in that, In S4, the formula for calculating the mean square error is: ； Where N is the number of predicted samples within a preset range; Let be the actual photovoltaic power of the i-th sample; Let be the predicted photovoltaic power of the i-th sample; The adaptive incremental learning rate update formula is: ； in Let t be the dynamic learning rate. The initial learning rate is within the preset initial range; The decay coefficient for a single iteration is within the preset decay range; This is an incremental training compensation term used to offset data drift errors during long-term model operation; the rolling update has a preset maximum number of updates, and the number of steps in a single incremental training session does not exceed the preset upper limit.

5. The photovoltaic power rolling prediction method integrating meteorological early warning and cloud cluster remote sensing according to claim 1, characterized in that, In S1, the multi-source data acquisition methods are as follows: meteorological data is acquired in real time through the official meteorological interface, and the data update delay does not exceed the preset delay threshold; satellite cloud image data is collected according to the preset acquisition cycle and the cloud movement vector parameters are obtained by inversion; historical power and real-time operating power data are retrieved from the photovoltaic power station monitoring system, and the integrity of the whole station data is not lower than the preset integrity threshold.

6. The photovoltaic power rolling prediction method integrating meteorological early warning and cloud cluster remote sensing according to claim 1, characterized in that, After constructing the long and short scale datasets, S1 performs a preprocessing process on the data, including outlier removal and normalization. Min-Max normalization is used to normalize the regular features to a preset numerical range. Separate weighted normalization is performed on the radiation intensity mutation samples at the second preset time granularity to strengthen the weight of short-term mutation features.

7. The photovoltaic power rolling forecasting method integrating meteorological early warning and cloud cluster remote sensing according to claim 1, characterized in that, In S4, the rolling incremental optimization adopts a deployment architecture that coordinates the edge and the cloud: the latest measured power plant operation data and meteorological data are collected by the edge and uploaded to the cloud. After the cloud triggers the incremental training mechanism, it completes the model parameter update and sends the updated model parameters to the edge, which then performs subsequent power prediction.

8. A photovoltaic power rolling forecasting system integrating meteorological early warning and cloud cluster remote sensing, used to implement the method described in any one of claims 1-7, characterized in that, It includes a data acquisition module, a satellite cloud image analysis module, a feature correction module, a fusion prediction module, and an edge-cloud collaborative rolling update module; The data acquisition module's output is connected to the satellite cloud image analysis module's input and the feature correction module's input, respectively. The satellite cloud image analysis module's output is connected to the feature correction module's input, the feature correction module's output is connected to the fusion prediction module's input, the fusion prediction module's output is connected to the edge-cloud collaborative rolling update module's input, and the edge-cloud collaborative rolling update module's parameter output is connected to the fusion prediction module's parameter update. The data acquisition module is used to collect conventional power plant operation data at a first preset time granularity and radiation intensity mutation data at a second preset time granularity, and simultaneously acquire meteorological early warning data and satellite cloud image remote sensing data. The satellite cloud image analysis module is used to invert cloud cluster motion vectors and cloud coverage parameters based on satellite cloud image data, and to quantify cloud cluster movement and occlusion characteristics. The feature correction module is used to calculate the impact weight of meteorological warnings and the gradient of sudden changes in irradiance, and to dynamically correct the meteorological and power fusion feature vector. The fusion prediction module is used to carry an Informer encoder-multiscale LSTM fusion model and output the initial predicted power of photovoltaic power with weights. The edge-cloud collaborative rolling update module is used to monitor prediction errors in real time, and after triggering a preset error threshold, it performs incremental training and adaptive learning rate updates to achieve model anti-drift rolling optimization.

9. The photovoltaic power rolling forecasting system integrating meteorological early warning and cloud cluster remote sensing according to claim 8, characterized in that, The feature correction module calculates the warning impact weight according to the preset weight formula, and generates the corrected fusion feature vector according to the preset feature correction formula; in the fusion prediction module, the Informer encoder is a multi-layer structure, configured with multiple attention heads and including a special attention head for short-term irradiance intensity mutation; the multi-scale LSTM is a multi-layer structure, corresponding to multiple time-series scales of different durations; the dual model outputs are weighted by fusion weights to obtain the initial prediction power.

10. The photovoltaic power rolling forecasting system integrating meteorological early warning and cloud cluster remote sensing according to claim 8, characterized in that, The edge-cloud collaborative rolling update module uses a preset mean squared error as the incremental training trigger threshold and adopts an adaptive learning rate formula to perform parameter updates; the rolling update is set with a preset maximum number of times, and a preset upper limit for the number of steps in a single incremental training is set; the total delay of the system's overall end-to-end prediction output does not exceed a preset delay threshold.

Images