A total radiation prediction data revision method based on satellite inversion and ground monitoring

Abstract

The application discloses a total radiation prediction data revision method based on satellite inversion and ground monitoring, and relates to the field of prediction data revision.The method comprises the following steps: acquiring prediction data and inversion data corresponding to total radiation of a target photovoltaic station, and simultaneously collecting monitoring data of the total radiation of the target photovoltaic station; establishing a mapping model based on the inversion data, correcting the monitoring data by using the mapping model, and outputting a total radiation monitoring correction value according to the correction result; taking the total radiation monitoring correction value as input and combining meteorological parameters in a target period to establish a prediction data revision model, and outputting a revision result of the prediction data of the target photovoltaic station in combination with a verification rule.The application is based on satellite inversion and ground monitoring, realizes revision of numerical mode total radiation prediction data, and thus provides complete total radiation time series prediction data for a photovoltaic power station, and improves short-term prediction accuracy of power generation of the photovoltaic power station.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of forecast data revision, in particular to a total radiation forecast data revision method based on satellite inversion and ground monitoring. BACKGROUND

[0002] The fluctuation, randomness and intermittence caused by the seasonal and day-night output changes of photovoltaic power generation itself will have certain impact on the safe and stable operation of the power grid when it is connected in large scale. The most important influencing factor of photovoltaic power generation is total radiation, and the accuracy of total radiation forecast determines the accuracy of future photovoltaic power generation power prediction. However, the total radiation forecast based on numerical mode or statistical method always has errors.

[0003] Therefore, revision of numerical weather mode total radiation forecast data is of great significance to improve new energy consumption level, improve power grid dispatching capacity and efficiency.

[0004] At present, there is no effective solution to the problems in the related art. SUMMARY

[0005] In view of the problems in the related art, the present application provides a total radiation forecast data revision method based on satellite inversion and ground monitoring to overcome the above technical problems existing in the prior art.

[0006] To this end, the specific technical solutions adopted by the present application are as follows:

[0007] In a first aspect, the present application provides a total radiation forecast data revision method based on satellite inversion and ground monitoring, which comprises:

[0008] Obtaining the forecast data and inversion data corresponding to the total radiation of the target photovoltaic site, and collecting the monitoring data of the total radiation in the target photovoltaic site;

[0009] Establishing a mapping model based on the inversion data, and correcting the monitoring data using the mapping model, and outputting the total radiation monitoring correction value according to the correction result;

[0010] Taking the total radiation monitoring correction value as input combined with the meteorological parameters in the target period, establishing a forecast data revision model, and outputting the revision result of the forecast data of the target photovoltaic site combined with the verification rule.

[0011] Preferably, obtaining the forecast data and inversion data corresponding to the total radiation of the target photovoltaic site, and collecting the monitoring data of the total radiation in the target photovoltaic site comprises:

[0012] According to the time and spatial resolution screening rule, the total radiation forecast time series and the total radiation inversion time series of the target photovoltaic site are obtained as the forecast data and inversion data of the total radiation of the target photovoltaic site.

[0013] Based on the time resolution rule and the time span requirement, monitoring data of a total radiation station in a target photovoltaic station is acquired.

[0014] Preferably, a mapping model is established based on the inversion data, and the monitoring data is corrected using the mapping model, and a total radiation monitoring correction value is output according to a correction result.

[0015] The neural network model trained and fused with satellite multi-spectrum is used as the mapping model to realize mapping of satellite spectrum and ground irradiance.

[0016] A reference map is generated based on the inversion data, and the inversion data is subjected to interpolation alignment and abnormal time elimination processing, and a time delay embedding is constructed according to a processing result, and an embedding point set is output.

[0017] A persistent map is output by constructing a point set complex with the embedding point set, and compared with the reference map to verify the accuracy of the inversion data after the abnormal time elimination processing, so as to realize preprocessing of the inversion data.

[0018] The cross-source consistency of the preprocessed inversion data is judged, and the mapping model is migrated to the target photovoltaic station to generate an attention map according to a judgment result, and accurate inversion data is obtained.

[0019] The monitoring data correction weight is screened according to a traversal rule based on the accurate inversion data, and a total radiation monitoring correction value is output to correct the monitoring data.

[0020] Preferably, a persistent map is output by constructing a point set complex with the embedding point set, and compared with the reference map to verify the accuracy of the inversion data after the abnormal time elimination processing, so as to realize preprocessing of the inversion data, including:

[0021] The embedding point set in the target window is taken as a geometric space scatter point, the persistent map is generated by connecting the geometric space scatter point, and the persistent map is compared with the reference map, and the matching mode of the embedding point set is judged according to a comparison result.

[0022] Whether the inversion data after the abnormal time elimination processing has a structure destroyed condition is judged based on the matching mode.

[0023] If there is no structure destroyed condition, the inversion data after the abnormal time elimination processing is obtained, and if there is a structure destroyed condition, the inversion data eliminated by the abnormal time is subjected to a greedy backfilling processing, and the inversion data is reconstructed until there is no structure destroyed condition.

[0024] The inversion data exceeding the theoretical range in the inversion data after the abnormal time elimination processing is marked, and the inversion data exceeding the theoretical range is filtered by combining the deviation from the mean value, so that the preprocessing of the inversion data is realized.

[0025] Preferably, the cross-source consistency of the pre-processed inversion data is judged, the mapping model is migrated to the target photovoltaic site to generate an attention map according to the judgment result, and accurate inversion data are obtained, which include:

[0026] The data source credibility score of the prediction data, the pre-processed inversion data and the monitoring data is calculated, and the variation coefficient and the root mean square error ranking of the contents of the target sliding window are obtained to determine the statistical stability item and the historical performance item;

[0027] The physical consistency item is analyzed based on the self-consistency of the prediction data and the instantaneous irradiance, and the weights of the statistical stability item, the historical performance item and the physical consistency item are adjusted to calculate the relative difference between each other;

[0028] If the relative difference is greater than the difference threshold, the data source with the highest data source credibility score is selected as the reference data, if the reference data violates the physical constraint, the data source with the data source credibility score meeting the screening rule is selected as the reference data again, and if the relative difference is less than or equal to the difference threshold, the reference data screening process is skipped;

[0029] The prediction data, the pre-processed inversion data and the monitoring data are input into the pre-trained discriminator, the loss value and the reference data weighted fusion are output, the generator loss is obtained, and the data source is optimized based on the generator loss;

[0030] The mapping model is migrated to the target photovoltaic site according to the optimization result, combined with the numerical boundary mode to generate the attention map, and the attention map is combined with the multi-modal attention fusion layer, and the accurate inversion data are output according to the fusion result.

[0031] Preferably, the monitoring data correction weight is screened according to the traversal rule based on the accurate inversion data, and the total radiation monitoring correction value is output to correct the monitoring data, which includes:

[0032] Four same values greater than zero, data with time less than zero and greater than the target threshold in the daytime are screened from the monitoring data, and the corresponding index value is recorded, and the thickness weight is calculated based on the inversion thickness;

[0033] The correction factor is calculated based on the thickness weight and the atmospheric stratification index, and the thickness weight is used as the error judgment condition correction weight coefficient, combined with the site pressure of the target photovoltaic site and the accurate inversion data to calculate the total radiation monitoring correction value.

[0034] Preferably, the total radiation monitoring correction value is used as input combined with the meteorological parameters in the target period to establish a prediction data correction model, and the correction result of the prediction data of the target photovoltaic site is output combined with the verification rule, which includes:

[0035] The revision result of the forecast data based on the total radiation monitoring correction value output is output, and the revision result is divided into training input and verification input, and meteorological parameters in a target period are obtained as auxiliary input;

[0036] The training input and the auxiliary input are used as the input of the reinforcement learning to establish a revision model, and the output of the revision model is verified by using the verification input, and the learning rate and the discount factor of the revision model are adjusted according to the output result, so that an optimized revision model is obtained;

[0037] The forecast data is input into the optimized revision model, and the revision result of the forecast data of the target photovoltaic station is output based on the principle of minimizing the mean square error before and after revision and in combination with the verification rule.

[0038] In a second aspect, the present application further provides a total radiation forecast data revision system based on satellite inversion and ground monitoring, which comprises:

[0039] A data acquisition module is configured to acquire the forecast data and the inversion data corresponding to the total radiation of a target photovoltaic station, and to collect monitoring data of the total radiation of the target photovoltaic station;

[0040] A monitoring correction value output module is configured to establish a mapping model based on the inversion data, to correct the monitoring data by using the mapping model, and to output the total radiation monitoring correction value based on the correction result;

[0041] A revision result output module is configured to acquire the total radiation monitoring correction value as input, to establish a forecast data revision model in combination with meteorological parameters in a target period, and to output the revision result of the forecast data of the target photovoltaic station in combination with a verification rule.

[0042] In a third aspect, the present application further provides an electronic device, which comprises a memory, a processor, and a computer program stored in the memory and executable on the processor, and the computer program is executed by the processor to implement the above method.

[0043] In a fourth aspect, the present application further provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement the above method.

[0044] The present application has the following beneficial effects:

[0045] The present application is based on satellite inversion and ground monitoring, and realizes revision of numerical mode total radiation forecast data. By establishing a numerical mode total radiation forecast revision model, an integrated learning model of satellite inversion, mode forecast and ground actual radiation data for short-term total irradiance forecast revision is obtained, so that complete total radiation time series forecast data is provided for photovoltaic power stations, and the short-term prediction accuracy of photovoltaic power station power generation is improved. BRIEF DESCRIPTION OF DRAWINGS

[0046] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed in the embodiments. Obviously, the drawings described below only constitute some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained from these drawings without creative labor.

[0047] Figure 1 is a flow chart of a total radiation prediction data revision method based on satellite inversion and ground monitoring according to an embodiment of the present application;

[0048] Figure 2 is a principle block diagram of a total radiation prediction data revision system based on satellite inversion and ground monitoring according to an embodiment of the present application;

[0049] Figure 3 is a structural schematic diagram of a hardware running environment related to the embodiment scheme of the present application;

[0050] Figure 4 is a flow chart of a total radiation prediction data revision method based on satellite inversion and ground monitoring according to an embodiment of the present application;

[0051] Figure 5 is an output flow chart of a total radiation monitoring correction value in a total radiation prediction data revision method based on satellite inversion and ground monitoring according to an embodiment of the present application;

[0052] Figure 6 is a comparison chart of the effect of revising the predicted irradiance on 2022.05.31 according to an embodiment of the present application;

[0053] Figure 7 is a comparison chart of the effect of revising the predicted irradiance on 2022.06.01 MLP according to an embodiment of the present application;

[0054] Figure 8 is a comparison chart of the effect of revising the predicted irradiance on 2022.05.31 SVR according to an embodiment of the present application;

[0055] Figure 9 is a comparison chart of the effect of revising the predicted irradiance on 2022.06.01 SVR according to an embodiment of the present application;

[0056] Figure 10 is a comparison chart of the effect of revising the predicted irradiance on 2022.05.31 XGBoost according to an embodiment of the present application;

[0057] Figure 11 is a comparison chart of the effect of revising the predicted irradiance on 2022.06.01 XGBoost according to an embodiment of the present application;

[0058] Figure 12 is a comparison chart of full-sky measured irradiance in different months according to an embodiment of the present application;

[0059] Figure 13 is one of comparison charts of measured irradiance in different weather types according to an embodiment of the present application;

[0060] Figure 14 is another of comparison charts of measured irradiance in different weather types according to an embodiment of the present application;

[0061] Figure 15 is a comparison chart of sudden change in irradiance that is not captured by a forecast according to an embodiment of the present application;

[0062] Figure 16 is a comparison chart of full-sky measured irradiance in different months according to an embodiment of the present application;

[0063] Figure 17 is a comparison chart of weather type forecast according to an embodiment of the present application. DETAILED DESCRIPTION

[0064] To further illustrate the embodiments, the present application provides accompanying drawings, which are part of the disclosure and mainly serve to illustrate the embodiments. With reference to the relevant description of the embodiments, those skilled in the art should be able to understand other possible implementations and advantages of the present application.

[0065] According to an embodiment of the present application, a total irradiance forecast data revision method based on satellite inversion and ground monitoring is provided.

[0066] The present application will be further described in conjunction with the accompanying drawings and specific embodiments. As shown in Figure 1 and Figure 4 , the total irradiance forecast data revision method based on satellite inversion and ground monitoring according to an embodiment of the present application includes:

[0067] Step S1, obtaining the forecast data and inversion data corresponding to the total irradiance of the target photovoltaic site, and collecting the monitoring data of the total irradiance in the target photovoltaic site.

[0068] In one embodiment, obtaining the forecast data and inversion data corresponding to the total irradiance of the target photovoltaic site, and collecting the monitoring data of the total irradiance in the target photovoltaic site includes: obtaining the total irradiance forecast time series and the total irradiance inversion time series of the target photovoltaic site as the forecast data and inversion data of the total irradiance of the target photovoltaic site according to the time and spatial resolution screening rules; and obtaining the monitoring data of the total irradiation station in the target photovoltaic site based on the time resolution rules and time span requirements.

[0069] In step S2, a mapping model is established based on the inversion data, the mapping model is used to correct the monitoring data, and a total radiation monitoring correction value is output according to a correction result.

[0070] As shown in FIG. 1, in one embodiment, the method comprises the following steps. Figure 5 As shown in FIG. 1, in one embodiment, the method comprises the following steps.

[0071] In one embodiment, the step of constructing a point set complex to output a persistent graph with the embedded point set and comparing the persistent graph with the reference graph to verify the accuracy of the inversion data after the abnormal time elimination processing to achieve the preprocessing of the inversion data comprises the following steps.

[0072] In one embodiment, judging the cross-source consistency of the preprocessed inversion data, migrating the mapping model to the target photovoltaic site to generate an attention map according to the judgment result to obtain accurate inversion data includes: calculating the data source credibility score of the forecast data, the preprocessed inversion data and the monitoring data, and obtaining the variation coefficient and root mean square error ranking of each type of data in the target sliding window content, and determining the statistical stability item and the historical performance item; based on the self-consistency of the forecast data and the instantaneous irradiance, the physical consistency item is analyzed, and the weights of the statistical stability item, the historical performance item and the physical consistency item are adjusted, and the relative difference between each other is calculated; if the relative difference is greater than the difference threshold, the data source with the highest data source credibility score is selected as the reference data, if the reference data violates the physical constraint, the data source with the data source credibility score meeting the screening rule is selected as the reference data again, if the relative difference is less than or equal to the difference threshold, the reference data screening process is skipped; the forecast data, the preprocessed inversion data and the monitoring data are input into the pre-trained discriminator, and the loss value and the weighted fusion of the reference data are output, the generator loss is obtained, and the data source is optimized based on the generator loss; according to the optimization result, the mapping model is migrated to the target photovoltaic site to generate an attention map combined with a numerical boundary mode, and the attention map is combined with a multi-modal attention fusion layer, and the accurate inversion data is output according to the fusion result.

[0073] In one embodiment, based on the accurate inversion data, the monitoring data correction weight is screened according to the traversal rule, the total radiation monitoring correction value is output, and the monitoring data is corrected, including: selecting four same values greater than zero, data with time less than zero and greater than the target threshold from the monitoring data, and recording the corresponding index value, and calculating the thickness weight based on the inversion thickness; based on the thickness weight and the atmospheric layer index, the correction factor is calculated, and the thickness weight is taken as the error judgment condition correction weight coefficient, combined with the site pressure of the target photovoltaic site and the accurate inversion data to calculate the total radiation monitoring correction value.

[0074] Step S3, taking the total radiation monitoring correction value as input combined with the meteorological parameters in the target period, establishing a forecast data correction model, and outputting the correction result of the forecast data of the target photovoltaic site combined with the verification rule.

[0075] In one embodiment, the total radiation monitoring correction value is combined with the meteorological parameters in the target period as input to establish a forecast data correction model, and the corrected results of the forecast data of the target photovoltaic station are output in combination with the verification rule, including: based on the total radiation monitoring correction value, the revised results of the forecast data are output, and the revised results are divided into training input and verification input, and the meteorological parameters in the target period are obtained as auxiliary input; the training input and the auxiliary input are used as the input of the reinforcement learning to establish the correction model, and the output of the correction model is verified by using the verification input, and the learning rate and the discount factor of the correction model are adjusted according to the output results to obtain the optimized correction model; the forecast data are input into the optimized correction model, and the revised results of the forecast data of the target photovoltaic station are output in combination with the verification rule.

[0076] As shown in Figure 2 According to another embodiment of the present application, a total radiation forecast data correction system based on satellite inversion and ground monitoring is also provided, which comprises:

[0077] The data acquisition module 1 is used to acquire the forecast data and inversion data corresponding to the total radiation of the target photovoltaic station, and collect the monitoring data of the total radiation in the target photovoltaic station;

[0078] The monitoring correction value output module 2 is used to establish a mapping model based on the inversion data, correct the monitoring data by using the mapping model, and output the total radiation monitoring correction value according to the correction result;

[0079] The correction result output module 3 is used to combine the total radiation monitoring correction value with the meteorological parameters in the target period as input to establish a forecast data correction model, and output the corrected results of the forecast data of the target photovoltaic station in combination with the verification rule.

[0080] It should be explained that, based on satellite inversion and ground monitoring, the numerical mode total radiation forecast data is corrected in this embodiment, the satellite inversion, mode forecast and ground actual radiation data integrated learning model for short-term total irradiance forecast correction is obtained by establishing the numerical mode total radiation forecast correction model, so that complete total radiation time series forecast data is provided for the photovoltaic power station, and the short-term prediction accuracy of the photovoltaic power station is improved. In order to achieve the above purpose, the data correction method is as shown in Figure 4 The method is performed according to the following steps:

[0081] Step one, data collection and processing, collect and process WRF numerical mode forecast data, third generation meteorological stationary satellite Himawari-8 inversion data, and total radiation station monitoring data at the target photovoltaic station or nearby.

[0082] Specifically, WRF numerical model prediction data is collected to obtain total radiation prediction grid data with a time resolution of 15 minutes and a spatial resolution of 3 km, and 2m temperature, 2m humidity, ground pressure, 10m wind speed, low cloud amount, medium cloud amount, high cloud amount, 850hPa wind speed, and 850hPa temperature. The time span of the obtained data is more than one year, and the total radiation prediction time series WY i , i = 1, …, n (i.e. prediction data), i represents an index, and n represents the number of data.

[0083] The observation data of Japan's first third-generation meteorological geostationary satellite Himawari-8 publicly released by the Japan Meteorological Agency is downloaded to obtain actual total radiation inversion and spectral data with a time resolution of 15 minutes and a spatial resolution of 500 m. The time span of the downloaded data is one year (corresponding to the time of the WRF model prediction data), and the total radiation inversion time series WF i , i = 1, …, n (i.e. inversion data).

[0084] The total radiation station monitoring data at the target photovoltaic station or nearby is collected to obtain total radiation monitoring data with a time resolution of 15 minutes, and the data time span is one year, which is recorded as total radiation monitoring time series WJ i , i = 1, …, n (i.e. monitoring data).

[0085] Step two, monitoring data correction, using the satellite feature extraction network pre-trained by global multi-climate zone historical data, constructing a physically guided spatial attention module, the input of which includes boundary layer height data of numerical model, dynamically weighting the contribution of satellite and physical branch through channel attention, and finally establishing a correction model of inversion data and monitoring data to calculate the total radiation monitoring correction value.

[0086] It needs to be explained that since the photovoltaic station monitoring data may have data quality problems, the inversion data needs to be used to correct the monitoring data.

[0087] 1. Inversion data pre-training: using a deep neural network model pre-trained on other climate zones or a large number of historical stations, which integrates satellite multispectral / multichannel information as a mapping model, to realize the mapping of satellite spectrum and ground irradiance.

[0088] 2. Data preprocessing: including time alignment, spatial matching, data quality labeling and filtering, specifically including:

[0089] (1) Time alignment;

[0090] Interpolation alignment: if the data time resolution is inconsistent, linear interpolation or cubic spline interpolation is used to align the data to 15 minutes.

[0091] Sliding window average: for data with time deviation less than 5 minutes, take the mean of a sliding window of ±2.5 minutes centered at the target time point as the alignment value.

[0092] Outlier time rejection: mark data points with missing timestamps or obviously unreasonable, and exclude them in subsequent correction.

[0093] Where the anomaly raises verification includes:

[0094] The sequence is normalized to the irradiance index: k t = GHIt / GHI t cs ;

[0095] In the formula, GHI t is the monitoring irradiance (also written as y t ), GHI t cs is the clear sky irradiance, t represents the time interval, such as 0, 15, 30, etc. 1, 2, 3, etc.

[0096] Construct time delay embedding (Takens): y t =[k t ,k t-τ ,…,k t-(m-1)τ ] T ;

[0097] In the formula, m is the embedding dimension (usually 3-6, here 4), τ is the time delay step (here 2, i.e. 30 min interval).

[0098] The Vietoris-Rips complex is constructed with the embedded point set within the window W (here ±2 hours), and the persistent homology (H0, H1 focus) is obtained by increasing the scale ε.

[0099] Vietoris-Rips complex is to regard these points as scattered points in geometric space, and the original time line becomes a point cloud. Assuming a radius ε, if the distance between two points is less than ε, draw a side; three points form a triangle, and more points form higher-dimensional simplices. With the gradual increase of ε, a series of different connected relationships (topological structures) are obtained.

[0100] Calculate the persistent diagram D clean of the cleaned graph and the reference diagram D ref .

[0101] D ref is like the historical data without missing data, or the data before rejection. Compare the persistent diagram D clean with the reference diagram D refWe can check whether the outliers removal destroys the overall structure, e.g. the original diurnal cycle is invisible, measured by bottleneck distance:

[0102] ;

[0103] where γ is a matching between two point sets (i.e. D clean is matched to D ref ), is the coordinate difference between two points (the maximum dimension), is the distance of the worst pair in the matching, is the minimum scheme that selects the worst distance in all possible matching.

[0104] If d B >τ B , we determine that the structure is possibly destroyed, where τ B is the 95% quantile of the historical daily distribution.

[0105] Minimization reconstruction: for the removed set R, we do a greedy backfilling, trying to recall each point to see the drop of d B Δd B , and select the maximum one, iterate until d B ≤τ B or reach the upper limit K.

[0106] (2) Spatial matching;

[0107] Grid nearest neighbor matching: for the 3km grid data of the model prediction, we select the nearest grid to the station as the input.

[0108] Satellite pixel weighted fusion: for the 500m resolution satellite data, we take a 3x3 pixel area centered at the station, and give a Gaussian weight according to the distance between the pixel and the station, and then get the matching value by weighted average.

[0109] Topographic correction: if the elevation difference between the station and the matching grid is >100m, we introduce a topographic correction coefficient:

[0110] Correction coefficient = 1 + 0.01·(H site -H grid );

[0111] where H site , H grid are the elevations of the station and the grid.

[0112] (3) Data quality labeling and filtering;

[0113] Physical threshold filtering: label the data that is out of the theoretical range (nighttime irradiance >0 W / m 2, daytime irradiance > 1200 W / m 2 , can refer to relevant standards for implementation.

[0114] Statistical anomaly detection: label data points deviating from mean ± 3σ based on sliding Z-score (window = 24 hours).

[0115] 3. Cross-source consistency check:

[0116] If the difference between satellite, model, and station data is > 30%, further judgment is needed.

[0117] (1) Dynamic confidence score:

[0118] Calculate the confidence score C for each data source in real time. The weight distribution is based on the score rather than a fixed rule:

[0119] C source = w 物理 P source + w 统计· S source + w 历史 ·H source ;

[0120] The physical consistency term P includes:

[0121] Satellite: Check the spectral inversion residual (e.g., whether the difference between Himawari-8 channel 6 and channel 7 radiation is within the theoretical range);

[0122] Model: Verify the self-consistency of WRF output parameters;

[0123] Station: Detect whether the cosine relationship between instantaneous irradiance and solar elevation angle is reasonable.

[0124] The statistical stability term S includes:

[0125] The coefficient of variation CV = σ / μ of data within a sliding window (1 hour), S = e -CV , where σ represents the standard deviation, μ represents the arithmetic mean, and e represents the natural constant approximately 2.71828.

[0126] The historical performance term H includes:

[0127] The RMSE percentile ranking of the same type of weather in the past 30 days (the lower the ranking, the higher H).

[0128] Weight adjustment:

[0129] w 物理 = 0.6, w 统计 = 0.3, w 历史 = 0.1.

[0130] Rainy w 物理 = 0.3, w 统计 = 0.5, w 历史 = 0.2.

[0131] (2) Conflict detection and arbitration rules:

[0132] Three-source difference detection: Calculate the relative difference δ = |A-B| / max(A, B) between each other, A, B are specific numerical values of any two data sources to be compared, such as satellite, model, and site, if δ > 30%, do not directly take the site as the standard, but execute the following arbitration process:

[0133] Priority of credibility: Choose the highest data source as the reference. source

[0134] Physical constraint filtering: If the reference data violates the physical constraints (such as clear sky radiation exceeding the theoretical value), select the second highest credibility data source.

[0135] Adversarial verification generation: Input the three-source data into the pre-trained GAN discriminator, output the most likely real value G GAN , and weighted fusion with the reference data:

[0136] G final = 0.7·G benchmark + 0.3·G GAN ;

[0137] G benchmark represents the reference data or reference value, such as the site value.

[0138] (3) Adversarial training enhances GAN design;

[0139] Generator: Input conflicting three-source data, output real value estimation;

[0140] Discriminator: Input real historical data or generated data, judge authenticity and add physical rule loss:

[0141] L 物理 = ||Discriminator output - physical constraint compliance|| 2 ;

[0142] Training data: Construct three-source conflict scenarios (simulate satellite cloud blocking, model initial field error, site snow cover).

[0143] (4) Real-time feedback optimization;

[0144] Each arbitration result is stored in the database, and the H source historical performance score of each data source is dynamically updated.

[0145] ​If a data source fails arbitration for 3 consecutive times, trigger an alarm and reduce its default weight by 30%.

[0146] 4. Migrate the mapping model to the target site and generate an attention map based on the numerical model boundary layer height and other parameters using satellite spectra, numerical model boundary layer height, cloud water path, ground meteorological station relative humidity and wind speed data.

[0147] (1) Core parameter selection;

[0148] Select key meteorological parameters that affect photovoltaic monitoring data quality:

[0149] Boundary layer height (PBLH): determines atmospheric stability (low→stable, high→unstable);

[0150] Cloud water path (LWP / IWP): affects the penetration of satellite spectra;

[0151] Ground meteorological data: relative humidity (RH), wind speed (WS), used to correct weights.

[0152] (2) Attention weight calculation;

[0153] Adjust the weights of each data source according to different weather conditions:

[0154] Boundary layer height (PBLH):

[0155] PBLH<500m (stable)→reduce satellite weight (0.2), increase ground observation weight (0.8);

[0156] 500m≤PBLH≤1000m (neutral)→balance satellite (0.5) and ground data (0.5);

[0157] PBLH>1000m (unstable)→increase satellite weight (0.7), reduce ground weight (0.3).

[0158] Humidity (RH) correction:

[0159] RH>70%→satellite weight×0.5 (high humidity reduces satellite reliability);

[0160] RH<40%→satellite weight×1.2 (dry weather relies more on satellite).

[0161] Wind speed (WS) correction:

[0162] WS>5m / s→boundary layer disturbance enhanced, satellite weight×1.1;

[0163] WS<2m / s→stable boundary layer, ground observation weight×1.2.

[0164] (3) Final attention map generation, mapping the calculated weights to the numerical prediction grid;

[0165] Integrate the weight of each data source: satellite weight = PBLH weight x RH correction x WS correction;

[0166] Ground weight = 1 - satellite weight;

[0167] Then automatically balance the contribution of multi-source information according to the current weather conditions (such as reducing the weight of satellite spectrum in high humidity), as the input of the multi-modal attention fusion layer. This fusion layer needs to only unfreeze the last three layers of the network to avoid overfitting of small samples, and obtain accurate inversion data.

[0168] (1) Cloud phase correction;

[0169] If the satellite detects ice cloud (CloudPhase = Ice) and the ground temperature > 0℃, start optical thickness compensation: δ ice =k·IceWaterPath 0.7 ;

[0170] (2) When the relative humidity (RH) > 90%, force to use the physical branch as the main branch:

[0171] λ phys =min(1.0, λ phys ·(1+0.5·(RH0.9))).

[0172] 5、According to the following rules to traverse the monitoring data: greater than 0 for 4 consecutive same values; values less than 0 during the day; values greater than 1374, and record the i (index) values of these numbers.

[0173] 6、Calculate the atmospheric influence factor:

[0174] Cloud optical thickness weight: τ weight =exp(-τ sat / 15); τ sat is the satellite inversion cloud optical thickness;

[0175] Atmospheric stratification index: L index =(H 850 -H surface ) / 1000·(1+RH 700 / 100);

[0176] Aerosol correction term (using Himawari-8 aerosol product):

[0177] A corr =1-0.2·tanh((AOD-0.3) / 0.1);

[0178] 7、Error classification:

[0179] The specific details are shown in the classification table in Table 1:

[0180] Table 1: Classification Table

[0181]

[0182] Where δ=WF i- WJ i t represents the cloud movement time in minutes, AOD represents the atmospheric air quality retrieved from the satellite, and PM10 represents the PM10 concentration retrieved from the satellite.

[0183] 8. Multi-scale correction: WJ revised =(WF i / avg(WF)) a(T,P) ×avg(WF)·f 3D +ΔW;

[0184] (1) Dynamic index:

[0185] a(T,P)=0.6+0.2·sin(2π(DOY-80) / 365)-0.1·P site / 1013;

[0186] In the formula, DOY represents the year-to-date period, DOY = current date - January 1st + 1; P represents the station's atmospheric pressure in hPa.

[0187] (2) Correction factor: f 3D =τ weight ×L index ×A corr WJ revised To finalize the monitoring values.

[0188] Step 3: Forecast data correction. Remove forecast data errors and calculate parameters such as 3-hour pressure and temperature variation, cloud cover coefficient, and wind speed coefficient. Use the first 70% of the forecast data as input and the first 70% of the total radiation monitoring correction data as output. Establish a model using the last 30% of the two data as input and output, respectively, to calibrate the model.

[0189] Based on the correction model to remove errors, the input is changed to WY. i The output was changed to WJ. i-revised Ultimately, WY can be obtained. i-revised , take WY i-revised The first 70% of the data is used as training input, corresponding to WJ i-revised As the training output; take WY i-revised The last 30% of the data was used as validation input, corresponding to WJ. i-revised This serves as the verification output.

[0190] (3) Take the parameters such as pressure and temperature variation, cloud cover coefficient, and wind speed coefficient for 3 hours as input, as follows:

[0191] 3-hour transformer P b3 =P0-P -3 For example, the 3-hour pressure change at 03:00 is the air pressure value at 03:00 minus the air pressure value at 00:00;

[0192] 3-hour variable temperature T b3 =T0-T -3 For example, the 3-hour transformer at 03:00 is calculated by subtracting the temperature value at 00:00 from the temperature value at 03:00;

[0193] Cloud coverage coefficient Cloud=Cloud 高 +Cloud 中 *2+Cloud 低 *4;

[0194] Wind speed coefficient Wind=Wind 10m *(1+0.1*RH), where RH is the relative humidity;

[0195] Atmospheric circulation index: Taking the station as the center, the geopotential heights (H500 and H925) at 500 hPa and 925 hPa are obtained for all grid points with a radius of 100 km, and the average value at 500 hPa (H500, H925) is calculated. mean ), maximum value, minimum value, and the distance (d) of the maximum value from the station. max ), the minimum distance from the station (d) min ).

[0196] Atmospheric circulation index:

[0197] H mean / 1000×(1+d max / 200-d min / 200)×exp(-∣H925-H500∣ / 300).

[0198] Boundary layer index: The sea-level pressure layer and temperature (T) at 500 hPa are calculated for all grid points with a radius of 100 km centered on the station. 850 T surface ), 500 hPa U wind speed and V wind speed (u 850 v 850 ).

[0199] Boundary layer index = 1 / 2(u 2 850 +v 2 850 )+(T 850 -Ts urface ) / 10.

[0200] (4) Take the selected training input as the input of Q-Learning and the training output as the output of Q-Learning to establish a Q-Learning correction model (referred to as a correction model).

[0201] (5) Take the selected verification input as the input of the correction model and the verification output as the output, adjust the Q value, learning rate and discount factor according to the principle of minimizing the root mean square error (or reaching the average reward convergence times), and adjust the mapping model according to the average reward of the last K steps and the change rate of the verification set RMSE:

[0202] Reward function R meta = RMSE improve + λ · PhysicalScore;

[0203] wherein the accuracy improvement reward is RMSE improve = (RMSE t-1 - RMSE t ) / RMSE t-1 , the physical consistency reward is PhysicalScore = 1.0 - 0.3 · Clear Sky Radiation + 0.4 · Terrain Gradient + 0.3 · (1 - Atmospheric Pressure Correlation Coefficient Square), and finally the optimized Q-Learning correction model is obtained.

[0204] (6) Take the forecast data as the input of the optimized Q-Learning correction model, and take the minimum RMSE before and after correction as the principle, and cooperate with the following verification to obtain the total radiation forecast data after correction:

[0205] Clear sky index score, the total radiation value after correction needs to be within this range:

[0206] Lower threshold = 0.2 × G clearsky × (1 + 0.5 × Cloud Proportion);

[0207] Upper threshold = 1.3 × G clearsky × (1 - 0.2 × AOD);

[0208] Dynamic offset = ± 0.1 × Past 7-day same period RMSE;

[0209] Wherein, G clearsky is the clear sky radiation.

[0210] If the forecast value < (lower threshold - dynamic offset), or the forecast value > (upper threshold + dynamic offset), it is marked as abnormal and enters step B, otherwise it enters step C.

[0211] B step: adaptive sliding window score, which specifically includes: window size self-adjustment and error mode learning;

[0212] Window size self-adjustment process sunny day for 24 hours long window (capture daily cycle regularity); cloudy day for 6 hours short window (fast response to cloud changes); extreme weather for 1 hour micro window (real-time tracking of mutations);

[0213] Error mode learning is calculated within the window, error inertia = (current error - window error mean) / window error standard deviation, if | error inertia | > 2, it is determined to be abnormal, and the correction model fine-tuning is started, otherwise it directly enters step C.

[0214] Step C: Cloud-irradiance score S cloud = Corr(Cloud Covert , t, ΔGt);

[0215] S cloud is the score index (about close to -1 is better), Cloud Covert is the cloud amount at t.

[0216] As Figure 6 shown in the figure, for a certain photovoltaic power station, the data time interval is: 2021 June 2 0:00 to 2022 June 2 0:00, a total of 365 days, the time resolution is 5 minutes, 288 points per day, according to the training set, test set, prediction set ratio of 6:3:1 distribution data sample, the first 219 days, a total of 63072 data as training set, the following 110 days as test set, a total of 31680 data, the remaining 36 days a total of 10368 data as prediction set, taking the measured irradiance as the target function, the irradiance correction model is established, the RMSE value of the predicted irradiance of the prediction data set of 2022.05.31 is calculated from the original 192.57 to 169.07, reduced by 12.20% of the original, the effect diagram is shown in Figure 6 (2022.05.31 corrected predicted irradiance effect comparison chart), the RMSE value of the predicted irradiance of the prediction data set of 2022.06.01 is reduced from the original 190.99 to 171.24, reduced by 10.34% of the original, the effect diagram is shown in Figure 7 (2022.06.01MLP corrected predicted irradiance effect comparison chart).

[0217] At the same time, the same scheme is adopted, in which the correction algorithm is replaced by the corrected predicted irradiance effect comparison chart of SVR as shown in Figure 8 (2022.05.31SVR corrected predicted irradiance effect comparison chart) and Figure 9 (2022.06.31SVR corrected predicted irradiance effect comparison chart), in which the corrected predicted irradiance effect comparison chart of SVR is shown in Figure 10(2022.05.31XGBoost revised prediction irradiance effect comparison chart) and Figure 11 (2022.06.01XGBoost revised prediction irradiance effect comparison chart) shown, while the six basic model irradiance revision results comparison table is shown in Table 2:

[0218] Table 2: Comparison of irradiance revision results of basic models

[0219]

[0220] And the prediction bias of irradiance mainly includes the following reasons:

[0221] (1) Different seasons, months, and different sunny day monitoring irradiance, the maximum irradiance is different, such as Figure 12

[0222] (2) The prediction accuracy of irradiance under different weather types is significantly different, and the accuracy of sunny days is significantly higher than that of non-sunny days, Figure 13 Figure 14

[0223] (3) Due to the limitation of numerical prediction resolution, sudden changes cannot be captured in time, Figure 15

[0224] (4) The prediction fails to capture the afternoon climbing process, Figure 16 Figure 17

[0225] In addition, the present application also provides an electronic device. As shown in the structural diagram of the hardware running environment of the electronic device, Figure 3 The electronic device can include: a processor, such as a CPU, a memory, a user interface, a network interface, and a communication bus. Among them, the communication bus is used to realize the connection communication between components. The user interface can include a display screen, an input unit such as a keyboard, and an optional user interface can also include a standard wired interface, a wireless interface. The network interface can optionally include a standard wired interface, a wireless interface. The memory can be a high-speed RAM memory, or a stable memory (non-volatile memory), such as a magnetic disk memory. The memory can also be an optional storage device independent of the aforementioned processor.

[0226] Those skilled in the art can understand that Figure 3 The electronic device shown in does not constitute a limitation on the electronic device, and can include more or fewer components than the diagram, or combine certain components, or different component arrangements.

[0227] Figure 3 ​​​​​​As shown, the memory as a computer storage medium can include an operating system, a network communication module, a user interface module and a device management program. The operating system is a program for managing and controlling hardware and software resources of the electronic device, and supports the running of the electronic device and other software or programs. In Figure 3 In the electronic device shown, the user interface is mainly used for connecting the terminal and communicating data with the terminal, such as receiving user signaling data sent by the terminal; the network interface is mainly used for the background server and communicating data with the background server; the processor can be used to call the program stored in the memory and execute the steps of the method or system as described above.

[0228] In addition, the embodiments of the present application also propose a computer readable storage medium, and the computer readable storage medium stores a device management program. The device management program is executed by the processor to realize the steps of the method or system as described above.

[0229] The computer readable storage medium of the present application has basically the same implementation as the above-mentioned method or system, and will not be described here. In addition, in order to achieve the above-mentioned purpose, the present application also provides a computer program product, which comprises: a computer program, which is executed by the processor to realize the steps of the method or system as described above.

[0230] The above-mentioned is only the preferred embodiment of the present application, and does not limit the present application. Any modification, equivalent replacement, improvement, etc. within the spirit and principle of the present application should be included in the protection scope of the present application.

Claims

1. A total radiation forecast data revision method based on satellite inversion and ground monitoring, characterized in that, The method comprises: obtaining forecast data and inversion data corresponding to total radiation of a target photovoltaic station, and simultaneously collecting monitoring data of the total radiation in the target photovoltaic station; establishing a mapping model based on the inversion data, correcting the monitoring data by using the mapping model, and outputting a total radiation monitoring correction value according to a correction result; combining the total radiation monitoring correction value as input with meteorological parameters in a target time period to establish a forecast data correction model, and outputting a correction result of the forecast data of the target photovoltaic station in combination with a verification rule; the method of establishing a mapping model based on the inversion data and correcting the monitoring data by using the mapping model, and outputting a total radiation monitoring correction value according to a correction result comprises: using a neural network model trained and fused with satellite multispectral as the mapping model to realize the mapping of satellite spectrum and ground irradiance; generating a reference graph based on the inversion data, and performing interpolation alignment and abnormal time elimination processing on the inversion data, and constructing a time delay embedding according to a processing result, and outputting an embedding point set; constructing a point set complex by using the embedding point set to output a persistent graph, and comparing the persistent graph with the reference graph to verify the accuracy of the inversion data after the abnormal time elimination processing, so as to realize the preprocessing of the inversion data; judging the cross-source consistency of the preprocessed inversion data, and migrating the mapping model to the target photovoltaic station to generate an attention graph according to a judgment result, so as to obtain accurate inversion data; screening monitoring data correction weights according to a traversal rule based on the accurate inversion data, and outputting a total radiation monitoring correction value to correct the monitoring data.

2. The total radiation forecast data revision method based on satellite inversion and ground monitoring according to claim 1, characterized in that, the method of obtaining forecast data and inversion data corresponding to total radiation of a target photovoltaic station, and simultaneously collecting monitoring data of the total radiation in the target photovoltaic station comprises: obtaining total radiation forecast time series and total radiation inversion time series of the target photovoltaic station as forecast data and inversion data of the total radiation of the target photovoltaic station according to a time and spatial resolution screening rule; obtaining monitoring data of a total radiation station in the target photovoltaic station based on a time resolution rule and a time span requirement.

3. The method of claim 1, wherein the method is characterized by, the method of constructing a point set complex by using the embedding point set to output a persistent graph, and comparing the persistent graph with the reference graph to verify the accuracy of the inversion data after the abnormal time elimination processing, so as to realize the preprocessing of the inversion data comprises: taking the embedding point set in a target window as a geometric space scatter point, connecting the geometric space scatter point to generate a persistent graph, and comparing the persistent graph with the reference graph to judge a matching mode of the embedding point set according to a comparison result; judging whether the inversion data after the abnormal time elimination processing has a structure damaged or not based on the matching mode; if there is no structure damaged, the inversion data after the abnormal time elimination processing is obtained, and if there is a structure damaged, the inversion data eliminated abnormally is processed by a greedy backfilling, and the inversion data is reconstructed until there is no structure damaged; marking inversion data exceeding a theoretical range in the inversion data after the abnormal time elimination processing, and filtering the inversion data exceeding the theoretical range by using a mean deviation, so as to realize the preprocessing of the inversion data.

4. The total radiation forecast data revision method based on satellite inversion and ground monitoring according to claim 3, characterized in that, the method of judging the cross-source consistency of the preprocessed inversion data, and migrating the mapping model to the target photovoltaic station to generate an attention graph according to a judgment result, so as to obtain accurate inversion data comprises: The data source credibility score of the calculation prediction data, preprocessed inversion data and monitoring data is evaluated, and the variation coefficient and root mean square error ranking of each type of data in the target sliding window content are obtained to determine the statistical stability item and historical performance item; Based on the self-consistency of the prediction data and the instantaneous irradiance analysis physical consistency item, the weights of the statistical stability item, the historical performance item and the physical consistency item are adjusted, and the relative difference between each other is calculated; If the relative difference is greater than the difference threshold, the data source with the highest data source credibility score is selected as the reference data, if the reference data violates the physical constraint, the data source with the data source credibility score meeting the screening rule is selected as the reference data again, if the relative difference is less than or equal to the difference threshold, the reference data screening process is skipped; The prediction data, preprocessed inversion data and monitoring data are input into the pre-trained discriminator, and the loss value and the reference data weighted fusion are output to obtain the generator loss, and the data source is optimized based on the generator loss; According to the optimization result, the mapping model is migrated to the target photovoltaic station, the attention map is generated combined with the numerical boundary mode, and the attention map is combined with the multi-modal attention fusion layer, and the accurate inversion data is output according to the fusion result.

5. The total radiation forecast data revision method based on satellite inversion and ground monitoring according to claim 4, characterized in that, The total radiation monitoring correction value is output by screening four same values greater than zero, data with time less than zero and greater than the target threshold in the monitoring data, and recording the corresponding index value, and calculating the thickness weight based on the inversion thickness; The correction factor is calculated based on the thickness weight and the atmospheric stratification index, and the thickness weight is taken as the error judgment condition correction weight coefficient, and the total radiation monitoring correction value is calculated combined with the site pressure of the target photovoltaic station and the accurate inversion data. The total radiation monitoring correction value is taken as the input combined with the meteorological parameters in the target period to establish a prediction data correction model, and the correction result of the prediction data of the target photovoltaic station is output combined with the verification rule, which includes:

6. The method of claim 1, wherein the method is characterized by, The revision result of the prediction data is output based on the total radiation monitoring correction value, and the revision result is divided into training input and verification input, and the meteorological parameters in the target period are obtained as auxiliary input; The training input and auxiliary input are taken as the input of the reinforcement learning to establish the correction model, and the learning rate and discount factor of the correction model are adjusted according to the output result of the verification input to obtain the optimized correction model; The prediction data is taken as the input of the optimized correction model, and the minimum mean square error before and after correction is taken as the principle to output the correction result of the prediction data of the target photovoltaic station combined with the verification rule; The meteorological parameters in the target period include variable pressure temperature, traffic coefficient, wind speed coefficient, atmospheric circulation index and boundary layer index; The boundary layer index is calculated based on the temperature corresponding to the target pressure height and the wind speed in the east-west direction and the north-south direction in the target radius range centered on the target photovoltaic station; The atmospheric circulation index is calculated based on the geopotential height of all grid points in the target radius range centered on the target photovoltaic station. The system comprises:

7. A satellite inversion and ground monitoring based total radiation forecast data revision system for implementing the satellite inversion and ground monitoring based total radiation forecast data revision method of any one of claims 1-6, characterized by, ​ The data acquisition module is configured to acquire forecast data and inversion data corresponding to total radiation of a target photovoltaic station, and to collect monitoring data of the total radiation of the target photovoltaic station. The monitoring correction value output module is configured to establish a mapping model based on the inversion data, to correct the monitoring data by using the mapping model, and to output a total radiation monitoring correction value according to a correction result. The correction result output module is configured to combine the total radiation monitoring correction value as input with meteorological parameters in a target period, to establish a forecast data correction model, and to output a correction result of the forecast data of the target photovoltaic station in combination with a verification rule.

8. An electronic device, comprising: The electronic device includes a memory, a processor, and a computer program stored on the memory and executable on the processor, and the computer program, when executed by the processor, implements the steps of the total radiation forecast data correction method based on satellite inversion and ground monitoring according to any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer program is stored on the computer readable storage medium, and when executed by the processor, implements the steps of the total radiation forecast data correction method based on satellite inversion and ground monitoring according to any one of claims 1 to 6.

Images