An ultra-short-term photovoltaic power self-adaptive prediction method and system

Abstract

The application discloses a kind of ultra-short photovoltaic power adaptive prediction method and system, belong to new energy power prediction technical field;The application eliminates time series deviation by multi-source data space-time registration, extracts cloud cluster thickness, deformation, split fusion multidimensional features based on light-weight U-Net semantic segmentation technology, combined with cloud cluster motion vector to construct comprehensive cloud amount influence coefficient, build double-layer irradiation correction model to realize accurate correction of meteorological data;Photovoltaic initial power prediction is completed using CNN-BiLSTM-Attention-ConvLSTM four-dimensional fusion model;A three-level irradiation sudden drop grading early warning mechanism is constructed to realize early prediction and power correction under different cloud layer shielding conditions;Finally, combined with energy storage SOC state, source storage collaborative power compensation is completed, and the output grid available net output prediction value and energy storage control instruction are output;The application solves the technical problems of traditional technology cloud cluster feature mining deficiency and poor model generalization, significantly improves the prediction accuracy in complex scenarios.

Description

Technical Field

[0001] This invention belongs to the field of new energy power prediction technology, specifically relating to an ultra-short-term photovoltaic power adaptive prediction method and system. Background Technology

[0002] Ultra-short-term photovoltaic (PV) power forecasting is a core support for real-time grid dispatching, renewable energy consumption, and coordinated control of energy storage charging and discharging. Existing traditional forecasting schemes have many inherent technical defects, which severely restrict the accuracy of PV power forecasting and the adaptability of system dispatching under complex meteorological scenarios.

[0003] Existing technologies rely solely on static numerical values ​​of total cloud cover and low cloud cover for modeling, failing to quantify the dynamic deformation and evolution processes of cloud expansion, contraction, splitting, and merging. They also rely only on three-dimensional vector features of cloud movement speed, direction, and coverage, resulting in significant lags in predicting irregular, fragmented, and rapidly deforming clouds, thus hindering the prediction of the entire cloud cycle's evolution trend. Furthermore, existing technologies utilize multi-source data from weather radar, satellite APIs, and power plant SCADA systems, which exhibit 2-5 minute time-series misalignments and time lags. The lack of a dedicated time-series calibration mechanism directly leads to distorted calculations of the irradiance correction coefficient, significantly increasing prediction errors for cloudy and fragmented cloud scenarios.

[0004] Furthermore, traditional forecasting techniques only provide numerical predictions of photovoltaic power without integrating with power plant energy storage systems to build a collaborative control mechanism. After receiving the forecast data, the grid cannot synchronously match the energy storage charging and discharging strategies, resulting in a disconnect between forecasting and control. Traditional models use a single, fixed learning rate for iterative optimization, making them unsuitable for diverse weather scenarios such as clear skies, cloudy skies, strong shading, and cloud cover. They suffer from slow convergence, large model oscillations, and significant long-term accuracy degradation when facing data distribution drift. Simultaneously, existing irradiance drop early warning mechanisms only set a single trigger threshold, failing to differentiate between mild shading, moderate drops, and extreme full shading conditions, thus failing to provide a tiered decision-making basis for grid frequency regulation and refined energy storage control.

[0005] Conventional CNN, LSTM, and CNN-BiLSTM-Attention prediction models can only capture basic temporal features and simple spatial features, lacking the ability to specifically extract spatiotemporal deformation features of clouds. In complex scenarios such as fragmented cloud deformation, extreme cloud full coverage, weak light in the early morning and late evening, and heavy rain, the normalized mean square error of prediction is generally higher than 15%, which cannot meet the scheduling requirements of new power systems for high-precision, low-latency, strong robustness, and source-storage coordination of photovoltaic power prediction. Summary of the Invention

[0006] To address the problems mentioned in the background art, this invention provides an ultra-short-term photovoltaic power adaptive prediction method and system, which solves the problems of single cloud feature extraction dimension, time sequence misalignment of multi-source data, weak model generalization ability, lack of hierarchical adaptive optimization, lack of hierarchical early warning mechanism, and lack of energy storage collaborative regulation capability in the prior art.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] An adaptive prediction method for ultra-short-term photovoltaic power includes the following steps:

[0009] S1: Collect multi-source data including photovoltaic power data, irradiance, temperature, cloud cover parameters, cloud cluster motion vector parameters, cloud cluster coverage, satellite cloud image sequence, and energy storage operation data. Unify the time reference of multiple data sources through spatiotemporal registration interpolation compensation algorithm, perform interpolation alignment on time-series misaligned data, and output a time-series unbiased standardized dataset.

[0010] S2: Perform lightweight U-Net semantic segmentation on continuous multi-frame satellite cloud images in the standardized dataset to extract cloud thickness, cloud deformation coefficient, and cloud splitting and fusion identifier; construct a cloud deformation evolution prediction sub-model to output cloud deformation prediction coefficients for a preset duration; combine cloud motion vector parameters to calculate cloud motion influence coefficient and cloud deformation influence coefficient. The cloud-related parameters extracted and calculated in S2 together constitute cloud feature parameters.

[0011] S3: Combine cloud cover parameters, cloud movement influence coefficient, and cloud deformation influence coefficient to jointly calculate the comprehensive cloud cover influence coefficient; construct a two-layer irradiance correction model. The first layer performs basic two-factor irradiance correction based on the comprehensive cloud cover influence coefficient to obtain the primary corrected irradiance. The second layer introduces the cloud deformation prediction coefficient and cloud thickness threshold to perform secondary compensation correction on the primary corrected irradiance and outputs the final corrected irradiance.

[0012] S4: Construct a CNN-BiLSTM-Attention-ConvLSTM four-dimensional fusion prediction model, using the final corrected irradiance, temperature, cloud cover parameters, photovoltaic power data, cloud motion vector parameters, cloud feature parameters, cloud coverage, and energy storage state parameters extracted from energy storage operation data as input features, and output the initial predicted photovoltaic power.

[0013] S5: Access the latest multi-source data after registration according to the preset period, calculate the normalized mean square error; pre-divide multiple meteorological scenarios, match differentiated iterative optimization parameters for each scenario; when the normalized mean square error exceeds the preset error threshold, call the adaptive learning rate formula of the driving quantity term of the corresponding scenario to dynamically update the model parameters and complete the hierarchical rolling optimization.

[0014] S6: Set multi-level irradiance drop warning thresholds, determine the warning level based on cloud coverage and cloud motion vector parameters, output warning signals with different preset durations in advance, and lower the predicted power lower limit in layers.

[0015] S7: Collects real-time state of charge and rated charge / discharge power of energy storage, calculates energy storage compensation power based on early warning level, adds energy storage charge / discharge compensation to the initial predicted power of photovoltaic, and outputs the predicted net power output of the grid and real-time control instructions for energy storage.

[0016] An ultra-short-term photovoltaic power adaptive prediction system, wherein the system is configured with signals sequentially connected along the data stream:

[0017] Multi-source data acquisition and time-series registration layer; includes satellite cloud image receiving unit, meteorological radar access unit, photovoltaic SCADA acquisition unit, energy storage BMS acquisition unit, and spatiotemporal interpolation compensation submodule; used to acquire multi-source data and perform spatiotemporal registration and time delay compensation, and output standardized dataset;

[0018] A cloud semantic segmentation and deformation prediction layer; connected to the multi-source data acquisition and temporal registration layer, with built-in lightweight U-Net cloud image segmentation module, cloud deformation evolution prediction sub-module, and cloud multi-dimensional feature calculation unit; used to extract cloud multi-dimensional features, predict cloud deformation trends, and calculate various cloud influence coefficients;

[0019] A cloud amount-cloud cluster dual-layer irradiance correction layer; connected to the cloud cluster semantic segmentation and deformation prediction layer, including a basic irradiance calculation unit, a cloud cluster deformation secondary compensation correction unit, and a comprehensive cloud amount influence coefficient solution unit; used to calculate the comprehensive cloud amount influence coefficient and output the final corrected irradiance.

[0020] A four-dimensional fusion prediction layer; connected to the cloud amount-cloud cluster dual-layer irradiance correction layer signal, including a CNN module, a BiLSTM module, an Attention module, a ConvLSTM module, and a fusion power output unit; used to output the initial photovoltaic prediction power;

[0021] A hierarchical adaptive rolling optimization layer; connected to the signal of the four-dimensional fusion prediction layer, including a meteorological scene recognition unit, a scene-specific momentum learning rate calculation unit, and a model parameter iterative update unit; used to perform scene-specific adaptive rolling updates of model parameters;

[0022] A three-level irradiance drop warning layer; connected to the cloud semantic segmentation and deformation prediction layer and the four-dimensional fusion prediction layer respectively; includes a three-level threshold determination unit, a power lower limit hierarchical correction unit, and a warning signal output unit; used to determine the warning level and hierarchically correct the predicted power lower limit.

[0023] The energy storage collaborative compensation and control layer is connected to the three-level irradiance drop early warning layer and the four-dimensional fusion prediction layer respectively. It includes an energy storage SOC reading unit, a graded compensation power calculation unit, and an energy storage charging and discharging command issuing unit, which are used to output the predicted value of the available net power output of the power grid and the energy storage control command.

[0024] Compared with the prior art, the beneficial effects of the present invention are:

[0025] 1. This invention adopts a technical link that combines lightweight U-Net cloud image semantic segmentation with cloud deformation evolution prediction. It breaks through the limitations of traditional two-dimensional features that rely solely on cloud amount and motion vectors, and adds multi-dimensional features such as cloud thickness, deformation, and splitting fusion to achieve prediction of cloud evolution trends throughout the entire cycle.

[0026] 2. This invention adds a multi-source data spatiotemporal registration interpolation compensation mechanism, which completely solves the problem of time sequence misalignment of data from multiple devices such as satellites, radars, and power stations, eliminates the problem of radiation correction distortion caused by data time delay, and effectively improves the accuracy of basic data.

[0027] 3. This invention proposes a hierarchical momentum adaptive learning rate optimization strategy for meteorological scenarios, abandoning the traditional global unified parameter iteration mode and matching exclusive iteration parameters for different meteorological scenarios, thereby improving the model convergence speed, robustness and stability.

[0028] 4. This invention constructs a three-level graded early warning system for sudden irradiance drops, replacing the traditional single threshold early warning system. It accurately distinguishes cloud cover scenarios, matches differentiated power correction strategies and scheduling response mechanisms, and significantly improves the accuracy of power grid scheduling redundancy configuration.

[0029] 5. This invention adds a photovoltaic-energy storage coordinated power compensation prediction model, which realizes the integrated linkage of photovoltaic power prediction and energy storage charging and discharging regulation, outputs the net available power output of the grid and simultaneously issues energy storage regulation commands, filling the industry gap of traditional prediction technology that only predicts photovoltaic power and cannot link energy storage regulation, and adapts to the source-storage coordinated dispatch requirements of new power systems. Attached Figure Description

[0030] Figure 1 This is a schematic diagram illustrating the specific process of this application. Detailed Implementation

[0031] 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.

[0032] An adaptive prediction method for ultra-short-term photovoltaic power, such as Figure 1 As shown, it includes the following steps:

[0033] S1: Multi-source data synchronous acquisition and spatiotemporal registration time delay compensation preprocessing:

[0034] Historical power data, real-time photovoltaic power, real-time irradiance, real-time temperature, total cloud cover, low cloud cover, cloud movement speed, cloud movement direction, and cloud coverage are collected. Simultaneously, satellite cloud image sequences from the meteorological bureau and ground radar echo data are acquired. The raw data time granularity is 10 minutes. A spatiotemporal registration interpolation compensation algorithm unifies the time reference of the three data sources: satellite, radar, and power station SCADA. For misaligned data with delays of 0-5 minutes, piecewise linear interpolation is used to complete time-series alignment, outputting a time-series unbiased standardized dataset. Low cloud cover refers to the proportion of the sky obscured by low-level clouds with cloud base heights ranging from 0m to 2000m, as observed in meteorological observations of the geographical area where the photovoltaic power station is located over time periods; the unit of measurement is percentage.

[0035] S2: Lightweight U-Net Semantic Segmentation Cloud Multidimensional Feature Extraction and Cloud Deformation Evolution Prediction:

[0036] Lightweight U-Net semantic segmentation was performed on three consecutive frames of registered satellite cloud images to extract cloud thickness, cloud deformation coefficients, and cloud splitting / fusion identifiers. A cloud deformation evolution prediction sub-model was constructed to output the cloud deformation prediction coefficients for the next 10-20 minutes. The cloud deformation evolution prediction sub-model uses a TCN-GRU temporal network structure to perform inference calculations. First, the input sequence is constructed using the temporal deformation features of the three consecutive frames output by the lightweight U-Net. , The measured cloud deformation baseline coefficients at times t-2, t-1, and t are shown in sequence. The input sequence is subjected to TCN dilated convolution operation to extract global temporal variation features. The convolution operation formula is:

[0037] ;

[0038] The specific meanings of each symbol in the formula are as follows: Input sequence for the temporal characteristics of cloud deformation; These are the deep temporal features extracted by TCN dilated convolution; These are the weight parameters corresponding to the convolution kernel positions; The kernel size; The dilation coefficient is the dilation factor of the dilated convolution. These represent the temporal feature components corresponding to the delayed time points. The convolution result is fed into the GRU unit to iteratively update the hidden state. The GRU state update calculation formula is as follows:

[0039] ;

[0040] ;

[0041] ;

[0042] ;

[0043] The specific meanings of each symbol in the formula are as follows: The gate output value is updated at the current moment to control the proportion of historical feature information retained. Reset the gate output value at the current moment to discard invalid historical time series features; The previous GRU unit's hidden state stores historical cloud deformation time-series characteristics; The current candidate hidden state represents the latest temporal feature fusion result; The hidden features are finally updated for the current moment and used as the output features of the sub-model's time-series inference. These are the learnable weight matrices for the update gate, reset gate, and candidate hidden layer, respectively. Use the Sigmoid activation function; The matrix multiplication operator is used for element-wise multiplication.

[0044] Hide features in GRU output After mapping through the fully connected layer, normalization constraint operations are performed using the Sigmoid activation function, ultimately yielding a result constrained to... Cloud deformation prediction coefficients in the interval The calculation formula is:

[0045] ;

[0046] In the formula: This is the output layer weight matrix. For bias terms; The larger the value, the more severe the cloud deformation in the next 10-20 minutes. After this coefficient is calculated, the cloud movement influence coefficient and cloud deformation influence coefficient can be calculated by combining the cloud movement speed, direction, and coverage.

[0047] The formula for calculating the influence coefficient of cloud movement is:

[0048] ;

[0049] The formula for calculating the cloud deformation influence coefficient is:

[0050] ;

[0051] in, is the cloud deformation prediction coefficient, with a value range of [0,1]. For real-time cloud thickness, The threshold for cloud thickness; The cloud cluster splits and merges, with the merged state K=0 and the splitting state K=1.

[0052] Furthermore, lightweight U-Net satellite cloud image cloud cluster multi-dimensional feature extraction includes the following steps:

[0053] 1) Perform grayscale normalization and noise filtering preprocessing on satellite infrared cloud images to remove interference pixels from the land surface and ocean outside the cloud layer;

[0054] 2) The lightweight U-Net encoder uses 3-layer downsampling to compress the original image to 1 / 8 resolution, extracts the spatial contour features of clouds, and the decoder is paired with skip connections to restore the image resolution;

[0055] 3) Quantitatively output the core features of three types of cloud clusters: average cloud cluster thickness, cloud cluster contour deformation coefficient, and cloud cluster splitting and fusion identifier;

[0056] 4) Input three consecutive time-series cloud images into a lightweight ConvLSTM sub-model to predict the cloud deformation prediction coefficients for the next 10-20 minutes, and output them to the irradiation correction calculation link.

[0057] S3: Joint Calculation of Comprehensive Cloud Cover Impact Coefficient and Multi-Level Irradiance Correction:

[0058] The comprehensive cloud cover influence coefficient is calculated by combining the total cloud cover, low cloud cover, cloud cluster motion influence coefficient, and cloud cluster deformation influence coefficient. A two-layer irradiance correction model is constructed. The first layer is the basic two-factor irradiance correction to obtain the primary corrected irradiance. The second layer introduces the cloud cluster deformation prediction coefficient and cloud cluster thickness to complete the secondary compensation correction and output the final corrected irradiance.

[0059] The formula for calculating the overall cloud cover impact coefficient is as follows:

[0060] ;

[0061] The specific formula for double-layer irradiation correction is as follows:

[0062] First-level base irradiation correction:

[0063] ;

[0064] Second-layer cloud deformation secondary compensation correction:

[0065] ;

[0066] in, This is the historical average irradiance for the same period. This represents the maximum cloud coverage.

[0067] S4: Power Prediction Using a Four-Dimensional Fusion Model

[0068] A CNN-BiLSTM-Attention-ConvLSTM four-dimensional fusion prediction model is constructed, which takes multiple corrected features, including corrected irradiance, temperature, cloud cover, cloud cluster features, and energy storage status, as input and outputs the initial predicted power of photovoltaics.

[0069] Specifically, the input feature vector of the four-dimensional fusion model includes: corrected irradiance, temperature, total cloud cover, low cloud cover, historical power, power change rate, cloud movement speed, cloud movement direction, cloud coverage, cloud thickness, cloud deformation coefficient, cloud deformation prediction coefficient, energy storage SOC value, energy storage rated power, and level 3 irradiance drop warning flag.

[0070] Specifically, the input feature vectors of the four-dimensional fusion model are divided into four categories and are respectively connected to the corresponding sub-modules for computation:

[0071] The first type of spatial meteorological texture features are fed into the CNN module. The spatial meteorological texture features include: corrected irradiance, temperature, total cloud cover, low cloud cover, cloud coverage, cloud thickness, cloud deformation coefficient, and cloud deformation prediction coefficient. The CNN module extracts the spatial distribution features of the cloud and outputs the spatial prediction power PCNN.

[0072] The second type of photovoltaic time-series operating features are fed into the BiLSTM module. The photovoltaic time-series operating features include: historical power and power change rate. The BiLSTM module mines the time-series fluctuation pattern of photovoltaic power and outputs the time-series predicted power PBiLSTM.

[0073] The third type of warning status feature is sent to the Attention module. The warning status feature is a three-level irradiance sudden drop warning flag. The Attention module adaptively adjusts the feature weights according to the warning flags and outputs the weight-corrected predicted power PAtt.

[0074] The fourth type of cloud dynamics and energy storage coupling characteristics are fed into the ConvLSTM module. The cloud dynamics and energy storage coupling characteristics include: cloud movement speed, cloud movement direction, energy storage SOC value, and energy storage rated power. The ConvLSTM module couples the spatiotemporal evolution of the cloud with the energy storage operating condition output and couples the predicted power PConvLSTM.

[0075] After the four sub-modules output four predicted powers in parallel, the arithmetic mean of PBiLSTM, PAtt, and PConvLSTM is first calculated, and then the average value is weighted and fused with PCNN through the fusion weight λ to obtain the initial predicted power of photovoltaic.

[0076] The output formula of the four-dimensional fusion model is as follows:

[0077] ;

[0078] The fusion weight λ∈[0.20,0.60] is determined by 10-fold cross-validation, and the validation set accounts for 15% of the total weight.

[0079] The model's module parameters are configured as follows: The CNN module has 4 convolutional layers with a kernel size of 3×3, a stride of 1, uses max pooling, has a dropout coefficient of 0.2, and adds a global average pooling layer with an output dimension of 128; the BiLSTM module has 2 layers with a hidden dimension of 320, uses a bidirectional output structure, and has a dropout of 0.15; the Attention module uses the GELU activation function, sets the attention weight dimension to 128, and adds a 4-head scene-aware multi-head attention branch; the ConvLSTM module has 2 layers with a kernel size of 5×5, used to extract the spatiotemporal correlation and deformation features of cloud motion.

[0080] S5: Layered adaptive scrolling model optimization:

[0081] Every 10-15 minutes, the latest registered measured data and meteorological data are input, and the normalized mean square error (NMSE) is calculated. Four meteorological scenarios are pre-divided: sunny, cloudy, rapidly deforming fragmented clouds, and extreme full obscuration. Differentiated initial learning rates, decay coefficients, and momentum coefficients are matched according to the scenario. When NMSE > 2.8%, the adaptive learning rate formula of the momentum term for the corresponding scenario is called to dynamically update the model parameters, completing hierarchical rolling optimization. The maximum number of rolling updates is 20-35 times.

[0082] Specifically, the formula for normalized mean square error is:

[0083] ;

[0084] The formula for updating the learning rate of hierarchical adaptive momentum is:

[0085] ;

[0086] Where s is the weather scene number, the sunny scene corresponds to an initial learning rate of 0.0006, a decay coefficient of 0.05, and a momentum coefficient of 0.3; the cloudy scene corresponds to an initial learning rate of 0.002, a decay coefficient of 0.07, and a momentum coefficient of 0.35; the fragmented cloud deformation passing scene corresponds to an initial learning rate of 0.005, a decay coefficient of 0.09, and a momentum coefficient of 0.4; and the extreme total obscuration scene corresponds to an initial learning rate of 0.008, a decay coefficient of 0.10, and a momentum coefficient of 0.5.

[0087] The number of predicted samples N=96~128. The Min-Max+Z-Score joint normalization method was used to normalize the conventional features to the [0,1] interval. The energy storage SOC, cloud thickness, and deformation coefficient were preprocessed using quantile normalization.

[0088] The multi-source data acquisition and registration methods are as follows: cloud cover, temperature, cloud cluster motion vector, and satellite cloud image data are acquired in real time through the meteorological bureau's satellite cloud image API combined with ground radar, with a raw data delay of ≤2 minutes and a time sequence deviation of ≤10s after spatiotemporal registration and interpolation compensation; irradiance, photovoltaic power, energy storage SOC, and energy storage charging and discharging power data are collected synchronously through the photovoltaic power station SCADA system combined with the energy storage BMS system, with data integrity ≥99.8%.

[0089] S6: Three-level irradiance drop classification early warning and power lower limit stratified correction:

[0090] Three levels of early warning thresholds are set. Level 1, mild warning, is when the cloud coverage is 70% ≤ cloud coverage < 80% and the cloud vector points towards the field area. A warning is issued 10 minutes in advance and the lower limit of the predicted power is slightly lowered. Level 2, moderate warning, is when the cloud coverage is 80% ≤ cloud coverage < 90% and the cloud vector points towards the field area. A warning is issued 15 minutes in advance and the lower limit of the power is moderately corrected. Level 3, extreme warning, is when the cloud coverage is ≥ 90% and the cloud fully covers the field area. A warning is issued 20 minutes in advance and the lower limit of the predicted power is significantly lowered. At the same time, a forced discharge command flag for energy storage is output.

[0091] S7: Photovoltaic-energy storage synergistic power compensation correction:

[0092] The system collects the real-time state of charge (SOC) and rated charge / discharge power of the energy storage, constructs a compensation correction formula for energy storage based on the graded early warning level, adds the energy storage charge / discharge compensation amount to the initial photovoltaic power prediction, outputs the predicted value of the net available power output of the grid, and simultaneously generates real-time control commands for energy storage.

[0093] The formula for synergistic compensation correction between photovoltaics and energy storage is:

[0094] ;

[0095] ;

[0096] in, For energy storage compensation power, discharge is a positive value and charging is a negative value; the graded compensation coefficients are α1=0.2, α2=0.5, and α3=0.8.

[0097] As a prediction system for performing the above prediction method, the system includes, in sequence along the data stream:

[0098] 1) Multi-source data acquisition and time-series registration layer: includes satellite cloud image receiving unit, meteorological radar access unit, photovoltaic SCADA acquisition unit, energy storage BMS acquisition unit, and spatiotemporal interpolation compensation submodule;

[0099] 2) Cloud semantic segmentation and deformation prediction layer: Built-in lightweight U-Net cloud map segmentation module, cloud deformation evolution prediction sub-module, and cloud multi-dimensional feature calculation unit;

[0100] 3) Cloud amount-cloud cluster double-layer irradiance correction layer: basic irradiance calculation unit, cloud cluster deformation secondary compensation correction unit, and comprehensive cloud amount influence coefficient solution unit;

[0101] 4) Four-dimensional fusion prediction layer: CNN sub-module, multi-head attention BiLSTM sub-module, ConvLSTM cloud spatiotemporal feature extraction sub-module, fusion power output unit;

[0102] 5) Hierarchical adaptive rolling optimization layer: meteorological scene recognition unit, scene-specific momentum learning rate calculation unit, and model parameter iterative update unit;

[0103] 6) Three-level irradiance drop early warning layer: three-level threshold determination unit, power lower limit layer correction unit, and early warning signal output unit;

[0104] 7) Energy storage collaborative compensation and control layer: energy storage SOC reading unit, hierarchical compensation power calculation unit, energy storage charging and discharging command issuing unit;

[0105] The total delay of the system's end-to-end prediction and energy storage command output is ≤15 seconds.

[0106] The invention will be described below with reference to specific application examples:

[0107] This invention selects a 15MW distributed photovoltaic power station in East China as the verification object. The photovoltaic modules are monocrystalline PERC modules, the inverter model is Sungrow SG1500CX, and a 2MW / 4MWh lithium iron phosphate energy storage system is built in conjunction with it. The model training hardware environment is CPU Intel i9-13900K, GPU NVIDIA RTX4090, 64GB memory, operating system is Ubuntu 22.04, deep learning framework is TensorFlow 2.15 and Python 3.10, and cloud map feature extraction is deployed on an NVIDIA Jetson Xavier edge computing device. Full-condition operation data from April to September 2025 are selected, covering all typical scenarios such as sunny days, cloudy days, rapid movement of fragmented clouds, cloudy to sunny days, short-term strong shading, rapid cloud transit, cloud fragmentation and deformation, extreme rainstorm shading, and low light in the morning and evening. Fixed thresholds are uniformly set: maximum cloud movement speed Vmax=20m / s, maximum cloud coverage Smax=100%, and maximum cloud thickness Hmax=10.

[0108] Example 1: Verification of sunny day working conditions prediction;

[0109] Operating parameters: Total cloud cover C=1, low cloud cover Clow=0, cloud clusters are stationary with no deformation, historical power 14200kW, real-time irradiance 980W / m², historical average irradiance for the same period 960W / m², ambient temperature 31℃. After spatiotemporal registration, the data time series is unbiased, the cloud cluster motion influence coefficient and deformation influence coefficient are both 0, the comprehensive cloud cover influence coefficient is 0.33, and the irradiance intensity after double-layer correction is 969.5W / m². The model fusion weight is 0.35, the initial predicted power output of the four-dimensional fusion model is 14212kW, the actual measured power at the site is 14230kW, the calculated NMSE is 1.72%, no model parameter update is triggered, and after 25 consecutive rolling updates, the error stabilizes at 1.6%~2.3%, demonstrating excellent prediction accuracy under clear weather conditions.

[0110] Example 2: Verification of slow cloud cluster movement under cloudy conditions;

[0111] Operating parameters: Total cloud cover C=6, low cloud cover Clow=4, cloud movement speed 3m / s, movement direction 45°, cloud cover 45%, historical power 8600kW, real-time irradiance 520W / m², historical average irradiance 580W / m², temperature 26℃. The calculated cloud movement influence coefficient is 0.078, with no deformation characteristics. The comprehensive cloud cover influence coefficient is 1.708, and the irradiance after double-layer correction is 589.3W / m². The model fusion weight is 0.45, the initial predicted power is 8618kW, the measured power is 8550kW, and the NMSE is 3.62%. Hierarchical adaptive optimization is triggered. After iteratively updating the model parameters, the average error stabilizes at 3.1%~3.7% after 15 consecutive rolling updates.

[0112] Example 3: Verification of the operation under the condition of rapid movement and shading by fragmented clouds;

[0113] Operating parameters: Total cloud cover C=9, low cloud cover Clow=8, cloud cluster movement speed 12m / s, movement direction 10°, coverage 85%, historical power 4200kW, real-time irradiance 210W / m², historical irradiance 350W / m², temperature 23℃. The calculated cloud cluster movement influence coefficient was 0.088, and the comprehensive cloud cover influence coefficient was 2.658. After double-layer deformation compensation correction, the irradiance accuracy was significantly improved. The initial predicted power error NMSE=5.92%, triggering a dedicated adaptive optimization strategy for fragmented cloud deformation scenarios, updating the learning rate to 0.0041. After optimization, the predicted power was 3745kW, with a deviation of only 25kW from the measured power, and the NMSE decreased to 2.58%.

[0114] Example 4: Verification of the cloud cluster movement during the cloudy to sunny transition;

[0115] Under initial high cloud cover conditions, with a total cloud cover of 8 and a low cloud cover of 6, the cloud clusters moved rapidly. After 30 minutes, the cloud clusters quickly moved out of the area, the total cloud cover dropped to 2, the low cloud cover dropped to 1, and the irradiance rapidly rebounded. The model of this invention can accurately track the dynamic changes in cloud cover, cloud cluster deformation, and motion state. It has 20 rolling updates and iterations throughout the process, with a maximum error of 3.9%, an average error of 3.0%, and no error exceeding the threshold, demonstrating excellent dynamic tracking capabilities under various operating conditions.

[0116] Example: Verification under short-term extreme full-coverage conditions;

[0117] Operating parameters: Total cloud cover 10, low cloud cover 10, cloud clusters fully cover the field area (100% coverage), triggering a Level 3 extreme irradiance drop warning, outputting a warning signal and a forced energy storage discharge command 20 minutes in advance. After correction for cloud thickness and deformation characteristics, the model's predicted power lower limit is corrected to 1650kW, the measured power is 1650kW, and the NMSE is 3.12%. High-precision prediction is maintained even under extreme shading conditions, and early energy storage regulation can be achieved.

[0118] Example 6: Verification of rapid splitting and deformation of cloud clusters;

[0119] Operating parameters: Total cloud cover 8, low cloud cover 7, cloud cluster movement speed 9 m / s, direction 60°, coverage 78%. U-Net semantic segmentation extracted cloud cluster thickness 7, deformation coefficient 0.75, splitting state identifier K=1, and deformation prediction value 0.72. The calculated cloud cluster movement influence coefficient was 0.1233, deformation influence coefficient was 0.0912, and comprehensive cloud cover influence coefficient was 2.5045. The model's initial predicted power was 4010 kW, triggering a level 1 mild warning. The energy storage compensation power was 400 kW, the net output prediction was 4410 kW, the measured net output was 4340 kW, and the NMSE was 2.41%. Compared with the traditional scheme without deformation correction, the prediction error was reduced by more than 40%.

[0120] Example 7: Verification of operation under extreme rainstorm and thick cloud passage conditions;

[0121] Under heavy rain conditions, the cloud thickness reaches the upper limit of the threshold, with a cloud coverage of 88%, triggering a level-two moderate warning. Through the dual-layer irradiance deformation compensation and correction mechanism of this invention, the irradiance prediction distortion caused by thick cloud cover is effectively offset. The model's predicted NMSE is reduced from 3.87% in the traditional scheme to 2.66%, adapting to the prediction needs of extreme weather conditions.

[0122] Example 8: Verification under low light and cloudy conditions in the early morning and late evening;

[0123] In low-irradiance, low-light environments during the early morning and late evening, the layered optimization parameters for sunny scenes of this invention effectively suppress the prediction drift problem of traditional models, and the prediction error is stable at ≤2.2% throughout the day, achieving high-precision prediction coverage of photovoltaic power stations throughout the day.

[0124] 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. An adaptive prediction method for ultra-short-term photovoltaic power, characterized in that, Includes the following steps: S1: Collect multi-source data including photovoltaic power data, irradiance, temperature, cloud cover parameters, cloud cluster motion vector parameters, cloud cluster coverage, satellite cloud image sequence, and energy storage operation data. Unify the time reference of multiple data sources through spatiotemporal registration interpolation compensation algorithm, perform interpolation alignment on time-series misaligned data, and output a time-series unbiased standardized dataset. S2: Perform lightweight U-Net semantic segmentation on continuous multi-frame satellite cloud images in the standardized dataset to extract cloud thickness, cloud deformation coefficient, and cloud splitting and fusion identifier; construct a cloud deformation evolution prediction sub-model to output cloud deformation prediction coefficients for a preset duration; combine cloud motion vector parameters to calculate cloud motion influence coefficient and cloud deformation influence coefficient. The cloud-related parameters extracted and calculated in S2 together constitute cloud feature parameters. S3: Combine cloud cover parameters, cloud cluster motion influence coefficient, and cloud cluster deformation influence coefficient to jointly calculate the comprehensive cloud cover influence coefficient; construct a two-layer irradiance correction model. The first layer performs basic two-factor irradiance correction based on the comprehensive cloud cover influence coefficient to obtain the primary corrected irradiance. The second layer introduces cloud deformation prediction coefficient and cloud thickness threshold to perform secondary compensation correction on the primary corrected irradiance, and outputs the final corrected irradiance. S4: Construct a CNN-BiLSTM-Attention-ConvLSTM four-dimensional fusion prediction model, using the final corrected irradiance, temperature, cloud cover parameters, photovoltaic power data, cloud motion vector parameters, cloud feature parameters, cloud coverage, and energy storage state parameters extracted from energy storage operation data as input features, and output the initial predicted photovoltaic power. S5: Access the latest multi-source data after registration according to the preset cycle, calculate the normalized mean square error; pre-divide multiple meteorological scenarios, and match differentiated iterative optimization parameters for each scenario; When the normalized mean square error exceeds the preset error threshold, the adaptive learning rate formula of the driving term for the corresponding scenario is called to dynamically update the model parameters and complete the hierarchical rolling optimization. S6: Set multi-level irradiance drop warning thresholds, determine the warning level based on cloud coverage and cloud motion vector parameters, output warning signals with different preset durations in advance, and lower the predicted power lower limit in layers. S7: Collects real-time state of charge and rated charge / discharge power of energy storage, calculates energy storage compensation power based on early warning level, adds energy storage charge / discharge compensation to the initial predicted power of photovoltaic, and outputs the predicted net power output of the grid and real-time control instructions for energy storage.

2. The ultra-short-term photovoltaic power adaptive prediction method according to claim 1, characterized in that, In S2, the formula for calculating the cloud cluster motion influence coefficient is: ； in, V is the cloud movement influence coefficient, and V is the cloud movement velocity. max The maximum cloud movement speed, The angle of direction of cloud movement; The formula for calculating the cloud deformation influence coefficient is: ； in, The cloud deformation influence coefficient. H represents the cloud deformation prediction coefficient, with values ​​falling within a preset normalization range; H is the real-time cloud thickness. K is the preset cloud thickness threshold; K is the cloud splitting and fusion indicator, which is 0 in the fusion state and 1 in the splitting state. In S3, the formula for calculating the comprehensive cloud cover impact coefficient is: ； in, The overall cloud cover impact coefficient is given by C, where C represents the total cloud cover in the cloud cover parameter. This refers to low cloud cover in the cloud cover parameter.

3. The ultra-short-term photovoltaic power adaptive prediction method according to claim 2, characterized in that, In S3, the specific formula for the double-layer irradiation correction is as follows: First-level base irradiation correction: ； Second-layer cloud deformation secondary compensation correction: ； in, For primary correction irradiation, For the final corrected irradiance, G represents the real-time irradiance intensity. The historical average irradiance for the same period is represented by S, where S is the cloud cover. This is the preset maximum cloud coverage.

4. The ultra-short-term photovoltaic power adaptive prediction method according to claim 3, characterized in that, In S4, the input feature vector of the four-dimensional fusion model includes corrected irradiance, temperature, total cloud cover, low cloud cover, historical power and power change rate in photovoltaic power data, cloud movement speed, cloud movement direction, cloud coverage, cloud thickness, cloud deformation coefficient, cloud deformation prediction coefficient, energy storage SOC value, energy storage rated power, and multi-level irradiance drop warning flag. The output formula of the four-dimensional fusion model is: ； in, This represents the initial predicted power of the photovoltaic system. To preset the fusion weights, This refers to the power prediction value output by the CNN module in the model. The power prediction value output by the BiLSTM module in the model. This refers to the power prediction value output by the Attention module in the model. This represents the power prediction value output by the ConvLSTM module in the model.

5. The ultra-short-term photovoltaic power adaptive prediction method according to claim 4, characterized in that, In S5, the formula for normalized mean square error is: ； in, For normalized mean square error, To predict the sample size, Let be the measured power of the i-th sample. Let be the predicted power of the i-th sample. This refers to the rated maximum power of the photovoltaic power station. The formula for updating the learning rate of hierarchical adaptive momentum is: ； in, Let be the learning rate for the t-th iteration in the s-th meteorological scenario. Let be the preset initial learning rate corresponding to the s-th type of meteorological scenario. The preset attenuation coefficient is the one corresponding to the s-th type of weather scenario. This is the preset momentum coefficient corresponding to the s-th type of weather scenario. Let be the momentum term for the t-th iteration.

6. The ultra-short-term photovoltaic power adaptive prediction method according to claim 5, characterized in that, In S5, the multi-level irradiance drop warning threshold is divided into three levels: A Level 1 mild warning corresponds to the first preset coverage area and the cloud vector points to the field area. It provides an early warning for the first preset duration and lowers the lower limit of the predicted power by the first magnitude. The Level 2 moderate warning corresponds to the second preset coverage range and the cloud vector points to the field area. It provides an early warning for the second preset duration and corrects the lower limit of power with the second amplitude. The Level 3 extreme warning corresponds to the third preset coverage range and the area fully covered by the cloud cluster. It provides an early warning for the third preset duration and lowers the predicted power limit by the third amplitude, while simultaneously outputting a forced discharge command flag for energy storage. The coverage thresholds of the first preset coverage interval, the second preset coverage interval, and the third preset coverage interval increase progressively, the first preset duration, the second preset duration, and the third preset duration increase progressively, and the first amplitude, the second amplitude, and the third amplitude increase progressively.

7. The ultra-short-term photovoltaic power adaptive prediction method according to claim 6, characterized in that, The specific calculations for S7 are as follows: ； ； in, This represents the predicted net available power output of the power grid. For energy storage compensation power, discharge is a positive value and charging is a negative value. The rated charge and discharge power for energy storage, , , This is the graded compensation coefficient.

8. An ultra-short-term photovoltaic power adaptive prediction system, used to implement the method according to any one of claims 1-7, characterized in that, The system sequentially connects and configures signals along the data stream: Multi-source data acquisition and time-series registration layer; includes satellite cloud image receiving unit, meteorological radar access unit, photovoltaic SCADA acquisition unit, energy storage BMS acquisition unit, and spatiotemporal interpolation compensation submodule; used to acquire multi-source data and perform spatiotemporal registration and time delay compensation, and output standardized dataset; A cloud semantic segmentation and deformation prediction layer; connected to the multi-source data acquisition and temporal registration layer, and includes a lightweight U-Net cloud image segmentation module, a cloud deformation evolution prediction sub-module, and a cloud multi-dimensional feature calculation unit; Used to extract multidimensional features of cloud clusters, predict cloud cluster deformation trends, and calculate the influence coefficients of various cloud clusters; A cloud amount-cloud cluster dual-layer irradiance correction layer; connected to the cloud cluster semantic segmentation and deformation prediction layer, including a basic irradiance calculation unit, a cloud cluster deformation secondary compensation correction unit, and a comprehensive cloud amount influence coefficient solution unit; Used to calculate the comprehensive cloud cover impact coefficient and output the final corrected irradiance; A four-dimensional fusion prediction layer; connected to the cloud amount-cloud cluster dual-layer irradiance correction layer signal, including a CNN module, a BiLSTM module, an Attention module, a ConvLSTM module, and a fusion power output unit; Used to output the initial predicted power of photovoltaic power; A hierarchical adaptive rolling optimization layer; connected to the four-dimensional fusion prediction layer, including a meteorological scene recognition unit, a scene-specific momentum learning rate calculation unit, and a model parameter iterative update unit; Used for adaptive rolling updates of model parameters based on different scenarios; A three-level radiation drop early warning layer; connected to the cloud semantic segmentation and deformation prediction layer and the four-dimensional fusion prediction layer respectively, including a three-level threshold determination unit, a power lower limit hierarchical correction unit, and an early warning signal output unit; Used to determine the warning level and stratify the lower limit of predicted power; The energy storage collaborative compensation and control layer is connected to the three-level irradiance drop early warning layer and the four-dimensional fusion prediction layer respectively. It includes an energy storage SOC reading unit, a graded compensation power calculation unit, and an energy storage charging and discharging command issuing unit, which are used to output the predicted value of the available net power output of the power grid and the energy storage control command.

9. The ultra-short-term photovoltaic power adaptive prediction system according to claim 8, characterized in that, In the multi-source data acquisition and time-series registration layer, cloud cover, temperature, cloud cluster motion vectors, and satellite cloud image data are acquired in real time through meteorological satellite data interfaces combined with ground radar. The original data delay does not exceed a preset delay threshold, and the time-series deviation after spatiotemporal registration interpolation compensation does not exceed a preset time-series deviation threshold. Irradiance, photovoltaic power, energy storage SOC, and energy storage charge and discharge power data are collected synchronously through the photovoltaic power station SCADA system combined with the energy storage BMS system, and the data integrity is not lower than the preset integrity index.

10. The ultra-short-term photovoltaic power adaptive prediction system according to claim 8, characterized in that, In the four-dimensional fusion prediction layer, the CNN module has 4 convolutional layers with a kernel size of 3×3 and a stride of 1. It uses max pooling with a dropout coefficient of 0.2 and adds a global average pooling layer, resulting in an output dimension of 128. The BiLSTM module has 2 layers with a hidden dimension of 320 and a bidirectional output structure with a dropout of 0.

15. The Attention module uses the GELU activation function, sets the attention weight dimension to 128, and adds a 4-head scene perception multi-head attention branch. The ConvLSTM module consists of two layers with 5×5 kernels, used to extract the spatiotemporal correlation and deformation features of cloud motion.

Images