Method and apparatus for dni prediction of a solar thermal power plant

Abstract

This disclosure provides a method and apparatus for predicting the DNI (Distribution-Influenced National) of a concentrated solar power (CSP) plant, relating to the field of prediction technology. The method includes: acquiring monitoring data of DNI prediction influencing factors for a first time period and forecast data of DNI prediction influencing factors for a second time period, centered on the target CSP, within a predetermined spatial range; preprocessing the acquired monitoring and forecast data; constructing model input features for predicting the DNI at the i-th time point within the prediction time period based on the preprocessed data; generating a sequence of predicted DNI points within the prediction time period based on the model input features and a pre-trained DNI prediction model; generating a prediction uncertainty interval based on the DNI point prediction value sequence; and outputting the final result of the DNI prediction for the target CSP. Based on this scheme, the accuracy of dynamic prediction of the CSP's DNI can be improved.

Description

Technical Field

[0001] This disclosure relates to the field of prediction technology, and in particular to a method and apparatus for predicting the DNI (Digital Incidence Niche) of a solar thermal power plant. Background Technology

[0002] Direct Normal Irradiance (DNI) is a core parameter for tower solar thermal power plants. Short-term (e.g., 0-6 hours, 10-minute resolution) DNI prediction is crucial for power generation planning, thermal storage management, and grid stability.

[0003] Typically, DNI's routine forecasts are based on a single data source, such as ground-based all-sky imager cloud cover data or a single digital weather forecast. The forecasting method is also relatively simple, such as mostly based on multiple linear regression and considering radiative transfer attenuation factor models.

[0004] However, DNI is highly volatile, and existing DNI prediction methods rely on a single modeling approach, resulting in predictions that frequently deviate from actual observations. Therefore, they cannot accurately predict short-term DNI. There is an urgent need for a method that can accurately predict short-term DNI to provide technical support for tower solar thermal power plants. Summary of the Invention

[0005] To overcome the problems existing in related technologies, this disclosure provides a method and apparatus for predicting the short-term dynamic NAND flash memory (DNI) of a solar thermal power plant, which can improve the accuracy of dynamic prediction of the short-term DNI of the power plant.

[0006] According to a first aspect of the present disclosure, a method for predicting the Dispersion Influence (DNI) of a solar thermal power plant is provided. The method includes: acquiring monitoring data of a preset DNI influence factor within a predetermined spatial range centered on the target power plant, and forecast data of a predicted DNI influence factor within a second time period, specifically within a first time period preceding the current time. The monitoring data includes monitored DNI values ​​and monitored values ​​of downward shortwave solar radiation provided by a monitoring satellite. The predicted DNI influence factor includes at least one of the following: aerosol optical thickness, PM2.5 concentration, and scattering coefficient from aerosol forecast data; and geopotential vorticity, cloud cover, temperature, humidity, and downward shortwave solar radiation from high-resolution numerical weather prediction data. The second time period... The process includes the predicted time period of the target power plant's DNI and a pre-set time period before the predicted time period; preprocessing the acquired monitoring and forecast data; constructing model input features for the DNI prediction at the i-th time point within the predicted time period based on the preprocessed data, where i is a positive integer; the model input features include at least one of the following: DNI lag features and DNI prediction influencing factors: original features, derived features, and time series features; generating a sequence of predicted DNI points within the predicted time period based on the model input features and the pre-trained DNI prediction model; generating a prediction uncertainty interval based on the DNI point predicted value sequence, and outputting the final result of the target power plant's DNI prediction.

[0007] Optionally, the lag feature indicates the DNI value at the (i-1)th time point when predicting the i-th time point. If i=1, the DNI value includes the DNI monitoring value within the first time period. If i>1, the DNI value includes at least the DNI predicted value before the i-th time point. The original feature indicates the original data of each DNI prediction impact factor. The derived feature includes the interaction term between a DNI prediction impact factor and other DNI prediction impact factors. The time series feature indicates the changing trend of the prediction impact factor.

[0008] Optionally, after obtaining the monitoring data of the DNI predicted impact factor in the first time period before the current time and the forecast data of the DNI predicted impact factor in the second time period, the method further includes: determining the target weather type corresponding to the target power plant at the current time and calling the DNI predicted target sub-model matched with the target weather type.

[0009] Optionally, a DNI prediction sub-model library is loaded, which includes pre-trained DNI prediction sub-models matching various weather types; if the weather type at the current time is the target weather type, the DNI prediction target sub-model corresponding to the target weather type is selected from the DNI prediction sub-model library; based on the model input features and the DNI prediction target sub-model, a sequence of DNI point prediction values ​​within the prediction time period is generated sequentially.

[0010] Optionally, the acquired target monitoring data and the forecast data of the DNI prediction impact factor are preprocessed, including at least one of the following: removing invalid values ​​from the target monitoring data and correcting outliers in the target monitoring data; aligning the timestamps of data from different sources for time synchronization; interpolating non-first resolution data to first resolution data, where the first resolution is the preset DNI prediction resolution; and standardizing the acquired data based on pre-trained normalization parameters.

[0011] Optionally, before generating the DNI point prediction value sequence within the prediction time period based on the model input features and the pre-trained DNI prediction model, the method further includes: acquiring raw data for training the DNI prediction model of the target power plant, the raw data including at least one of the following: ground monitoring data of the target power plant, reanalysis dataset, aerosol forecast data, shortwave radiation product data provided by satellite, and high-resolution numerical forecast data; the ground monitoring data includes DNI monitoring data and meteorological monitoring data; preprocessing the raw data to obtain preprocessed training data; determining the DNI prediction influencing factors based on the preprocessed training data; extracting the DNI prediction influencing factor data for each grid point within a preset size grid centered on the target power plant from the preprocessed training data; constructing model input feature data based on the extracted DNI prediction influencing factor data; and training the model based on the model input feature data and the training model to obtain the DNI prediction model.

[0012] Optionally, based on the preprocessed training data, the correlation between each factor and DNI is determined, and DNI prediction impact factors are selected based on the correlation; with DNI as the target, a regression model between each DNI prediction impact factor and DNI is established, the regression coefficient of each DNI prediction impact factor is determined, and DNI prediction impact factors used to train the DNI prediction model are selected based on the regression coefficient of each DNI prediction impact factor.

[0013] Optionally, based on the feature parameters, DNI lag features are generated according to the original features of the extracted first DNI predicted impact factor, and derived features and time series features corresponding to the first DNI predicted impact factor are created; the model input feature data for this training is generated according to the data of the original features, derived features, time series features, and DNI lag features of each DNI predicted impact factor; wherein, the first DNI predicted impact factor is any one of the extracted DNI predicted impact factors; the DNI lag feature indicates the DNI value before the i-th time point used to generate the DNI lag feature for the j-th training round at the i-th time point, where j is a positive integer. The derived features include the interaction term of the first DNI predicted impact factor and the second DNI predicted impact factor; the second DNI predicted impact factor is any one of the extracted DNI predicted impact factors that is different from the first DNI predicted impact factor; the time series features indicate the changing trend of the first DNI predicted impact factor; the feature parameters indicate the time range of the original features, the range of DNI values ​​before the i-th time point used to generate the DNI lag feature for the j-th training round for the i-th time point, the range of the second DNI predicted impact factor included in the derived features, and the time range corresponding to the time series features; the values ​​of the feature parameters are updated after each training round.

[0014] Optionally, based on the model input feature data and the training model, training can be performed separately for each weather scenario to obtain a DNI prediction sub-model for each weather scenario.

[0015] Optionally, based on the model input data corresponding to the target power station, a set of models for training is obtained; each model in the model set is trained separately, and the model with the highest prediction performance is selected as the training model for matching the target power station. The DNI prediction model obtained by training the training model for matching the target power station is used as the final DNI prediction model for the target power station.

[0016] According to a second aspect of the present disclosure, an apparatus for predicting the direct normal irradiance (DNI) of a solar thermal power plant is provided, comprising: a data acquisition module, a data preprocessing module, a model input feature construction module, and a DNI prediction module; the data acquisition module is used to acquire, within a preset spatial range centered on the target power plant, monitoring data of the direct normal irradiance (DNI) prediction influence factor for a first time period prior to the current time and forecast data of the DNI prediction influence factor for a second time period; the monitoring data includes monitored DNI values ​​and monitored values ​​of downward solar shortwave radiation provided by monitoring satellites; the forecast data includes at least one of the following: aerosol optical thickness, PM2.5 concentration, and scattering coefficient from aerosol forecast data, and geopotential vorticity, cloud cover, temperature, humidity, and downward solar shortwave radiation from high-resolution numerical forecast data; the second time period includes the target power plant... The system comprises: a prediction time period for the DNI of the target power plant and a preset time period preceding the prediction time period; a data preprocessing module for preprocessing the acquired monitoring and forecast data; a model input feature construction module for constructing model input features for the DNI prediction at the i-th time point within the prediction time period based on the preprocessed data, where i is a positive integer; the model input features include at least one of the following: DNI lag features and DNI prediction influencing factors: original features, derived features, and time series features; a DNI prediction module for generating a sequence of predicted DNI points within the prediction time period based on the model input features and a pre-trained DNI prediction model; and an uncertainty analysis module for generating a prediction uncertainty interval based on the DNI point predicted value sequence and outputting the final result of the DNI prediction for the target power plant.

[0017] According to a third aspect of the present disclosure, a server is provided, comprising: a processor; and a memory storing computer-readable instructions that, when executed by the processor, implement the method for predicting the DNI (Digital Incidence Negative Energy) of a solar thermal power plant as described in the first aspect.

[0018] According to a fourth aspect of the present disclosure, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the method for predicting the DNI (Dielectric Noise Injection) of a solar thermal power plant as described in the first aspect.

[0019] The technical solutions provided by the embodiments of this disclosure may include the following beneficial effects:

[0020] In this embodiment, firstly, monitoring data of the DNI prediction impact factor for the first time period and forecast data of the DNI prediction impact factor for the second time period are acquired within a preset spatial range centered on the target power station. Then, the acquired monitoring and forecast data are preprocessed, and based on the preprocessed data, model input features for the DNI prediction at the i-th time point within the prediction time period are constructed. Next, based on the model input features and the pre-trained DNI prediction model, a sequence of predicted DNI points within the prediction time period is generated sequentially. Finally, based on the DNI point prediction value sequence, a prediction uncertainty interval is generated, and based on the DNI point prediction value sequence and the prediction uncertainty interval, the final result of the DNI prediction for the target power station is generated and output. Because it deeply integrates aerosol forecast data, shortwave radiation monitoring and forecast values, and multi-source data such as aerosol optical thickness, PM2.5 concentration, and scattering coefficient, it constructs a model input feature that includes at least one of the following: original features, derived features, and time series features, achieving multi-dimensional synergy and overcoming the limitations of a single data source. By introducing real-time aerosol forecast data and cloud cover, we can simultaneously analyze cloud dynamics and aerosol concentration fluctuations, dynamically quantify the attenuation effect of dust particles on solar radiation, and dynamically capture the fluctuation effect of DNI. By using spatial features and derived features to improve the adaptability of the prediction model to complex weather, we can systematically improve the prediction accuracy of short-term DNI.

[0021] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure. Attached Figure Description

[0022] Figure 1 This is a system architecture diagram of a DNI prediction system for a solar thermal power plant provided in an embodiment of this disclosure.

[0023] Figure 2 This is a schematic diagram of the training process of a DNI prediction model provided in an embodiment of this disclosure.

[0024] Figure 3 This is a schematic diagram illustrating the change in the percentage of missing DNI observation data over the months, as provided in an embodiment of this disclosure.

[0025] Figure 4 This is a schematic diagram illustrating the change in the percentage of missing DNI observation data over time, as provided in an embodiment of this disclosure.

[0026] Figure 5 This is a schematic diagram illustrating the correlation between meteorological elements and DNI as a function of the present disclosure.

[0027] Figure 6This is a schematic diagram illustrating the hourly variation of the correlation between meteorological elements and DNI, as provided in an embodiment of this disclosure.

[0028] Figure 7 A schematic diagram illustrating the variation of the correlation between various meteorological elements of ERA5 and DNI with the month, as provided in this embodiment of the disclosure.

[0029] Figure 8 (a) Schematic diagram of the impact of ERA5 meteorological elements on DNI provided in the embodiments of this disclosure. Figure 1 .

[0030] Figure 8 (b) Schematic diagram of the impact of ERA5 meteorological elements on DNI provided in the embodiments of this disclosure. Figure 2 .

[0031] Figure 9 This is a schematic diagram of the processing flow of a method for predicting the DNI (Digital Intake) of a solar thermal power plant, provided as an embodiment of this disclosure.

[0032] Figure 10 A hardware structure diagram of a server housing the device for DNI prediction in a solar thermal power plant provided in this embodiment of the disclosure.

[0033] Figure 11 A block diagram illustrating DNI prediction for a solar thermal power plant, provided as an embodiment of this disclosure. Detailed Implementation

[0034] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this disclosure. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this disclosure as detailed in the appended claims.

[0035] The terminology used in this disclosure is for the purpose of describing particular embodiments only and is not intended to be limiting of the disclosure. The singular forms “a,” “the,” and “the” as used in this disclosure and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any and all possible combinations of one or more of the associated listed items.

[0036] It should be understood that although the terms first, second, third, etc., may be used in this disclosure to describe various information, such information should not be limited to these terms. These terms are used only to distinguish information of the same type from one another. For example, without departing from the scope of this disclosure, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to determination."

[0037] The embodiments of this disclosure will now be described in detail.

[0038] Figure 1 A system architecture diagram of a DNI prediction system for a concentrated solar power plant provided in this disclosure embodiment is shown below. Figure 1 As shown, the DNI prediction system 100 includes: a solar thermal power plant 101, a server 102, a first database 103 of a weather forecasting center, a satellite 104, an Earth observation system 105, and a second database 106 of the weather forecasting center. The server 102 can acquire ground monitoring data from the solar thermal power plant 101, aerosol forecast data from the first database 103 of the weather forecasting center, shortwave radiation monitoring data provided by the satellite 104, reanalysis datasets provided by the Earth observation system 105, and high-resolution numerical weather prediction data from the second database 106 of the weather forecasting center. The server 102 trains a model using machine learning based on the acquired data from various sources to obtain a DNI prediction model. During real-time forecasting, the server predicts the short-term (e.g., within 0-6 hours) DNI of the solar thermal power plant based on the pre-trained DNI prediction model and the data and weights of meteorological elements determined during pre-training.

[0039] First, in the method for predicting the DNI (Distributed National Intake) of a concentrated solar power plant provided in this embodiment, the training process of the DNI prediction model is given first, such as... Figure 2 The diagram shown is a schematic diagram of the training process of a DNI prediction model for a concentrated solar power plant provided in an embodiment of this disclosure, including the following steps S201 to S206:

[0040] S201. Obtain the raw data used to train the DNI prediction model for the target power plant.

[0041] The raw data includes at least one of the following: ground monitoring data of the target power plant, reanalysis dataset, aerosol forecast data, shortwave radiation product data provided by satellite, and high-resolution numerical forecast data; the ground monitoring data includes DNI monitoring data and meteorological monitoring data.

[0042] (1) Ground monitoring data

[0043] Ground-based monitoring data includes DNI monitoring data and meteorological observation data from the solar thermal power plant over a historical period. The meteorological observation data from the solar thermal power plant includes temperature, humidity, wind speed, and air pressure. The temporal resolution of the ground-based observation data is 10 minutes.

[0044] (2) Reanalyze the dataset

[0045] Reanalysis datasets are globally consistent, long-term series datasets of atmospheric, oceanic, or land conditions generated by combining historical observational data with numerical models using data assimilation techniques. Different institutions offer reanalysis datasets with varying advantages.

[0046] For example, in this embodiment of the disclosure, the fifth-generation atmospheric reanalysis dataset (ECMWF Atmospheric Reanalysis 5, EAR5) from the European Centre for Medium-Range Weather Forecasts (ECMWF) can be used as the reanalysis data in this embodiment. EAR5 can provide high-resolution meteorological data for high-frequency weather analysis. EAR5 includes temperature, humidity, wind speed, precipitation, cloud cover, and surface radiation. EAR5 has a temporal resolution of 1 hour and a spatial resolution of 0.025°.

[0047] It should be noted that, in the embodiments disclosed herein, reanalysis data is used for training historical features.

[0048] (3) Aerosol forecast data

[0049] Aerosol forecast data are predictions of the concentration, distribution, and transport processes of aerosols in the atmosphere (such as PM2.5 (particulate matter ≤2.5 μm, suspended particulate matter with a diameter ≤2.5 micrometers), PM10 (suspended particulate matter with a diameter ≤10 micrometers), dust, sulfates, black carbon, etc.). The parameters and accuracy of aerosol forecast data provided by different institutions vary.

[0050] For example, aerosol forecast data can be selected from NASA's Goddard Earth Observing System Composition Forecast (GEOS-CF) data. GEOS-CF provides real-time aerosol and chemical composition forecast data. NASA GEOS-CF aerosol forecast data includes aerosol optical thickness, scattering coefficient, and PM2.5 concentration. NASA GEOS-CF aerosol forecast data has a temporal resolution of 1 hour and a spatial resolution of 0.05°.

[0051] Aerosol optical thickness is typically a parameter characterizing the concentration of aerosols in the atmosphere and affects the attenuation of solar radiation.

[0052] (4) Shortwave radiation product data provided by satellite

[0053] Understandably, many satellites can provide shortwave radiation products, such as the Fengyun-4 satellite and the Kuihua-8 satellite. Depending on the needs, different shortwave radiation products from satellites that can acquire data can be selected.

[0054] For example, shortwave radiometric data provided by the Himawari-8 satellite can be selected. The Himawari-8 data (denoted as H8) has a temporal resolution of 10 minutes and a spatial resolution of 0.05°.

[0055] (5) High-Resolution Numerical Weather Prediction (HRNWP)

[0056] High-resolution numerical weather prediction (HMR) data is typically a core tool in meteorological and climate research and operational forecasting. Its spatial resolution usually ranges from kilometers to hundreds of meters, providing highly detailed weather forecasts. It can include global high-resolution models and regional high-resolution models.

[0057] For example, the high-resolution numerical forecast data from the European Centre for Medium-Range Weather Forecasts (ECMWF High Resolution Forecast System, EC-HRES), which can provide regional high-resolution models, can be selected as the high-resolution numerical forecast data in this disclosure embodiment. EC-HRES includes 52 variables, among which are temperature, humidity, and geopotential vorticity. EC-HRES has a temporal resolution of 1 hour and a spatial resolution of 0.1°.

[0058] It should be noted that the forecast data used in the training process are historical forecast data released at the same time, and the monitoring data used are historical monitoring data from the same period.

[0059] S202. Preprocess the original data to obtain preprocessed training data.

[0060] By preprocessing the original data, we can obtain training data that is time-synchronized, resolution-synchronized, free of outliers, and normalized.

[0061] S203. Based on the preprocessed training data, confirm the DNI prediction impact factor.

[0062] It is understandable that in this step, the DNI prediction system of the solar thermal power plant selects influence features that are related to DNI prediction from the training data.

[0063] S204. Extract the data of DNI predicted impact factors at each grid point within a preset spatial range with the target power plant as the grid center and a preset size grid from the preprocessed data.

[0064] For example, data from each grid point within a 4×4 (40km×40km) grid area centered on a solar thermal power plant can be used to train a DNI prediction model.

[0065] Understandably, during training, the relationship between the preset grid size and DNI prediction can be learned, so that raw data can be extracted based on the learned grid size, thereby reducing the acquisition of redundant data in the prediction stage and using less data while ensuring prediction accuracy.

[0066] S205. Based on the extracted DNI prediction impact factor data, construct the model input feature data.

[0067] It should be noted that, in this embodiment of the disclosure, during model training, for each forecast time point within the forecast interval, in each round of training, the DNI value from the acquired ground monitoring data is used as one of the data inputs to the model for the first forecast time point. After training for the i-th forecast time point, the DNI value predicted at the i-th forecast time point is used as one of the data inputs to the model for the (i+1)-th forecast time point, where i is an integer greater than 1.

[0068] S206. Train the DNI prediction model based on the model input feature data and the training model.

[0069] Specifically, for each forecast time point in a forecast interval of different lengths, training is performed based on the model input feature data corresponding to each forecast time point.

[0070] The time length of the forecast interval and the spacing between adjacent forecast time points can be set according to requirements, and this embodiment does not impose specific limitations on them.

[0071] For example, a short-term DNI prediction model can be trained to predict DNI values ​​every 10 minutes within the next 6 hours.

[0072] It should be noted that the trained DNI prediction model can indicate the meteorological elements included in the DNI prediction impact factors required for predicting DNI, the input features generated for each prediction impact factor and the impact weight of each input feature, the time period of the monitoring data in the DNI prediction impact factors to be obtained during prediction, the time period of the forecast data in the obtained DNI prediction impact factors, and the time period of the input data obtained for each forecast, including the start position, length and end position.

[0073] It is understood that different target power plants can all be trained to obtain DNI prediction models that match each target power plant based on the above training method. The prediction interval lengths of different target power plants can be different or the same. The time intervals of DNI predictions within the prediction intervals of each power plant can be the same or different. All of these can be trained as needed. This disclosure does not make any specific limitations on this.

[0074] Based on this scheme, before predicting the DNI of a target power plant, relevant meteorological monitoring and forecast data can be acquired. After preprocessing the relevant ground monitoring data, reanalysis datasets, aerosol forecast data, satellite-provided shortwave radiation product data, and high-resolution numerical forecast data, the DNI prediction influencing factors related to DNI prediction can be determined. After determining the DNI prediction influencing factors, the data for the DNI prediction influencing factors at each grid point within a preset spatial range centered on the target power plant and with a preset grid size is extracted. Model input feature data is constructed, and a training model is selected for training, ultimately resulting in a learned DNI prediction model. Because the training process provided in this embodiment integrates multi-dimensional, high-precision forecast data and monitoring data as collaborative input features, it can accurately capture the atmospheric attenuation state, thereby learning the degree of influence of aerosol optical thickness and scattering effects on DNI. By introducing the spatial features corresponding to the grid, accurate and short-term dynamic DNI predictions can be performed for the target power plant. Especially in dusty or polluted environments, it can dynamically quantify the attenuation effect of dust particles and the dynamic shading of clouds, thereby improving the accuracy of the DNI prediction model.

[0075] Optionally, the data preprocessing in this embodiment of the disclosure may include at least one of the following:

[0076] 1-1: Remove invalid values ​​from the acquired monitoring data and correct outliers in the monitoring data.

[0077] It is understandable that the raw data may contain missing data for some time periods due to sensor failure, communication interruption or other technical problems, or the data values ​​may remain unchanged for multiple consecutive moments (rigid data).

[0078] For example, in the data preprocessing process of this disclosure embodiment, missing and rigid data in the original data can be statistically filtered using T-test confidence intervals. For instance, rigid data refers to data values ​​that remain unchanged for ten consecutive time points. Rigid segments can be removed, interpolated, or re-estimated based on a physical model to correct rigid data.

[0079] Figure 3 This diagram illustrates the variation of the percentage of missing DNI observation data over months, as provided in an embodiment of this disclosure. Figure 3 The diagram shows the amount of missing data for each month, as well as the percentage of missing data for each month relative to the total missing data. This diagram can help identify peak periods or characteristic patterns of missing data in specific months. DNI data shows a higher number of missing data points during winter (October to December), with December accounting for 46% of the total missing data (nearly 22% of the total for that month). This phenomenon is likely closely related to severe winter weather conditions and difficulties in equipment maintenance; therefore, the impact of missing data should be considered in winter analyses.

[0080] Figure 4 This is a schematic diagram illustrating the change in the percentage of missing DNI observation data over time, as provided in an embodiment of this disclosure. Figure 4 The diagram shows the amount of missing data at various times throughout the day (1 AM to 12 AM), and the percentage of missing data at each time point relative to the total missing data. This can be used to identify the temporal distribution characteristics of missing data and reveal the central trend or abnormal patterns of missing data within a specific time period. However, missing data at different times may have an uneven impact on the overall 24-hour DNI assessment. Therefore, this embodiment of the disclosure considers this time-dependent data missing data when conducting a comprehensive analysis of the overall 24-hour DNI to ensure the scientific validity and reliability of the conclusions.

[0081] 1-2: Align timestamps from different data sources for time synchronization.

[0082] Aligning the timestamps of all data sources ensures consistency of data from each source in the time dimension, which can avoid analytical bias or model failure caused by time misalignment.

[0083] For example, if H8 is delayed by 5 minutes, the H8 data that is 5 minutes behind can be processed 5 minutes ahead of time.

[0084] 1-3: Interpolate non-first-resolution data to first-resolution data.

[0085] The first resolution is the preset resolution predicted by DNI.

[0086] Since the original data from different sources obtained in this embodiment have different time resolutions, the original data with different time resolutions can be unified to the same resolution through interpolation.

[0087] For example, interpolating data with a time resolution of 1 hour into data with a time resolution of 10 minutes.

[0088] For example, data that is approximately or linearly varying, such as temperature and humidity at a 1-hour resolution, is interpolated to a 10-minute resolution using a linear interpolation method to align with time; data that varies non-linearly, such as radiation related to zenith radiation and extinction coefficient, is interpolated to a 10-minute resolution using cubic splines to align with time.

[0089] 1-4: Feature Normalization

[0090] The acquired data is standardized based on the normalization parameters.

[0091] Standardizing the acquired raw data and mapping it to the range [0,1] can eliminate the impact of differences in scale and distribution on model training by scaling data with different features to the same numerical range.

[0092] For example, the Min-Max normalization method can be used to normalize and map the data of all elements in ERA5, NASA GEOS-CF, EC-HRES, and H8 to the range [0, 1].

[0093] It should be noted that during the model training phase, the model can be updated by adjusting the normalization parameters to obtain the final trained model, which means that the normalization parameters are also trained. During the model application phase, in the data preprocessing process, the acquired raw data can be standardized based on the pre-trained normalization parameters.

[0094] It should be noted that in the disclosed embodiments, the above-mentioned preprocessing is performed on the acquired raw data in both the model training stage and the model application stage, and this disclosed embodiment does not specifically limit this.

[0095] Based on this approach, the acquired raw data can be preprocessed before model training (or prediction), allowing for faster and more accurate learning of the influence weight coefficients of each factor, thereby improving the model's convergence speed.

[0096] Optionally, in the method for predicting the DNI of a solar thermal power plant provided in this embodiment of the present disclosure, the above-mentioned S203 may specifically include the following S203a and S203b.

[0097] S203a. Based on the preprocessed training data, determine the correlation between each factor and DNI, and select the DNI prediction influencing factor based on the correlation.

[0098] Specifically, based on the preprocessed raw data, key predictive factors of DNI can be screened using correlation coefficients. Correlation coefficients can be calculated using Pearson and Spearman correlation coefficients to determine the correlation between each factor and DNI. The Pearson correlation coefficient is used to measure the linear correlation between two continuous variables. r The range of values ​​is [0, 1]. r= -1 indicates a perfectly negative linear correlation. r= 0 indicates wireless correlation. r= A value of 1 indicates a perfectly positive linear correlation. The Spearman correlation coefficient measures the monotonic correlation between two variables; it is calculated based on the order of the variables. ρ The range of values ​​is [0, 1]. ρ= -1 indicates that the trend is completely monotonically decreasing. ρ= 0 indicates no monotonic relationship. ρ= 1 indicates a completely monotonically increasing trend.

[0099] Specifically, based on the calculation methods of Pearson correlation coefficient and Spearman correlation coefficient, the standardized time series of each factor of ERA5 data, H8 data, and surface meteorological data of the solar thermal power plant were calculated for each historical time period, and the correlation coefficients were calculated with the surface DNI time series of the solar thermal power plant. The correlation coefficients were then sorted in descending order of absolute value to select the DNI prediction influencing factors.

[0100] Optionally, weather conditions can be categorized based on weather phenomena in weather forecasts, such as dust storms, sunny days, and cloudy days.

[0101] Furthermore, for different scenarios, the correlation coefficients between the standardized time series of each factor under different weather types and the ground-based DNI time series of the solar thermal power plant can be calculated. The correlation coefficients are then sorted according to their absolute values ​​to determine the set of DNI prediction influencing factors corresponding to each weather type.

[0102] Table 1 is an exemplary table showing the annual and seasonal correlations between meteorological elements and DNI provided in this embodiment of the present disclosure.

[0103] Table 1. Annual and seasonal correlations between meteorological elements and DNI

[0104]

[0105] Referring to Table 1, from a seasonal perspective, in the region corresponding to a certain power station, winter temperature showed the strongest correlation with DNI, while humidity showed the strongest correlation in other seasons. Air pressure and wind speed showed no or weak correlation with DNI. Temperature was positively correlated with DNI in all seasons, with the strongest correlation in summer (0.4527) and the weakest in autumn (0.2735). Humidity was negatively correlated with DNI, with the strongest correlation in summer (-0.4546) and the weakest in winter (-0.3783).

[0106] Figure 5 This diagram illustrates the correlation between meteorological elements and the DNI (Density National Interference Factor) over the months, as provided in this embodiment of the disclosure. Looking at different months, the overall correlation follows a pattern consistent with the seasons. Specifically, the correlation between temperature and DNI exceeds 0.5 in November 2023, December 2023, and January 2024, with the strongest correlation in December 2023 at 0.5597. The correlation between humidity and DNI is less than -0.5 in August 2023 and October 2023, with the strongest correlation in October 2023 at -0.5339.

[0107] Figure 6 This diagram illustrates the hourly variation of the correlation between meteorological elements and the DNI (Distance-Induced Noise) as provided in this embodiment of the disclosure. From the perspective of peak output times, the correlation between humidity and DNI shows a pattern of first increasing and then decreasing, with the strongest correlation around 11:00 AM. The correlation between temperature and DNI shows a trend of first increasing and then decreasing in the early morning and evening, with a weak correlation between temperature and DNI from 9:00 AM to 3:00 PM. The correlation between air pressure and DNI is weak at all times. The correlation between wind speed and DNI is negative at all times, and the degree of correlation is low, indicating a weak or no correlation.

[0108] Figure 7 This diagram illustrates the monthly variation of the correlation between various meteorological elements and the DNI in ERA5 provided in this embodiment of the disclosure. The diagram shows the trend of the correlation between different variables and the DNI over the months. The overall conclusions are similar to those for seasonal variations. The most relevant variables are geopotential vorticity, temperature, and humidity.

[0109] Ultimately, calculations show a strong negative correlation between humidity, aerosol optical thickness, PM2.5 concentration, and surface DNI. Temperature, geopotential vorticity, and satellite shortwave radiation are positively correlated with surface DNI. Aerosol forecast data, along with some thermodynamic (temperature, satellite shortwave radiation) and dynamic (geopotential vorticity) meteorological data, are crucial for DNI prediction. The impact of aerosols is even more pronounced during dust storms.

[0110] S203b. With DNI as the target, establish a regression model between each DNI predicted impact factor and DNI, determine the regression coefficient of each DNI predicted impact factor, and select the DNI predicted impact factor to train the DNI prediction model based on the regression coefficient of each DNI predicted impact factor.

[0111] It is understandable that regression models can determine the degree of influence of each factor on DNI.

[0112] Specifically, based on the DNI prediction impact factors screened out according to correlation, and with DNI as the target, a multiple regression equation is established between the screened DNI prediction impact factors and DNI. Then, the regression coefficient of each DNI prediction impact factor is calculated using a multiple regression calculation method.

[0113] If modeling is done by weather type, then for each scenario, a multiple regression equation can be established between the DNI prediction impact factor selected by the correlation screening under different weather types and DNI, and the regression coefficient of each DNI prediction impact factor under each weather type can be calculated.

[0114] Specifically, the influence of each DNI prediction factor on DNI can be determined based on the ranking of the absolute values ​​of the regression coefficients of each DNI prediction influencing factor. Finally, the influence of each factor on DNI determined based on the Pearson correlation coefficient, Spearman correlation coefficient, and regression coefficient is obtained. Key factors are selected and used as the DNI prediction influencing factors of the DNI prediction model.

[0115] Figure 8 (a) Schematic diagram of the impact of ERA5 meteorological elements on DNI provided in the embodiments of this disclosure. Figure 1 , Figure 8 (b) Schematic diagram of the impact of ERA5 meteorological elements on DNI provided in the embodiments of this disclosure. Figure 2 Among them, in Figure 8 In (a), represents the ERA5 meteorological elements whose absolute regression coefficients are greater than 80. Figure 8 In (b), the ERA5 meteorological elements have an absolute value of regression coefficient less than or equal to 80.

[0116] Based on the experimental results, the regression coefficients for aerosol optical thickness (DUEXTTAU) were determined to be 120.5, PM2.5 regression coefficient to be -95.2, humidity regression coefficient to be -109.59, temperature regression coefficient to be 726.33, and satellite shortwave radiation regression coefficient to be 450.8. These results were used to guide feature selection. The selected influencing factors for DNI prediction included: aerosol optical thickness, PM2.5 concentration, and scattering coefficient from NASA GEOS-CF; shortwave radiation from H8 data; and geopotential vorticity, cloud cover, temperature, and humidity from EC-HRES.

[0117] For example, after selecting the DNI prediction influencing factors, a 4×4 grid (40km×40km) centered on the solar thermal power plant is selected, and meteorological data of the DNI prediction influencing factors at all grid points within the 4×4 grid are extracted for model training. This allows the prediction model to enhance its perception of regional atmospheric conditions based on the spatial characteristics within the 4×4 grid.

[0118] Based on this approach, the correlation between each factor and DNI can be determined first, and DNI prediction impact factors can be selected based on the correlation. Then, based on the importance of each DNI prediction impact factor, a set of DNI prediction impact factors from various data sources can be selected for model training. Training the model using data corresponding to the selected set of DNI prediction impact factors can improve the accuracy and efficiency of learning.

[0119] Optionally, in the method for predicting the DNI of a solar thermal power plant provided in this embodiment of the present disclosure, the above-mentioned S205 may specifically include the execution of S205a and S205b as described below.

[0120] S205a. Based on the feature parameters, according to the original features of the extracted first DNI predicted impact factor, generate DNI lag features, and create the derived features and time series features corresponding to the first DNI predicted impact factor.

[0121] During training, the first DNI prediction impact factor is any one of the extracted DNI prediction impact factors; the DNI lag feature indicates the DNI value before the i-th time point used to generate the DNI lag feature for the j-th round of training at the i-th time point, where j is a positive integer.

[0122] For each DNI predicted impact factor, the derived features include the interaction term between the extracted first DNI predicted impact factor and the second DNI predicted impact factor. The second DNI predicted impact factor is any of the extracted predicted impact factors that differs from the first DNI predicted impact factor. The time series features include the changing trend of the first DNI predicted impact factor.

[0123] Furthermore, the feature parameters indicate the time range of the original features, the range of DNI values ​​before the i-th time point used to generate the DNI lag features in the j-th training round for the i-th time point, the range of the second DNI prediction influence factor included in the derived features, and the time range corresponding to the time series features; the values ​​of the feature parameters are updated after each training round.

[0124] For example, the values ​​of feature parameters can be updated after each training round based on the loss function, such as adjusting the feature parameters based on backpropagation of the validation set error.

[0125] It should be noted that during training, when training on a specific timestamp, historical monitoring and forecast data after that timestamp can be obtained. In subsequent applications, for the model prediction time, if the monitoring values ​​of the DNI impact factor before and after the model prediction time are available, model input features, such as solar shortwave radiation, can be generated based on these values. If forecast values ​​after the model prediction time are not available, model input features, such as DNI ground monitoring values, can be generated based on monitoring values ​​before the model prediction time.

[0126] For example, the characteristic parameters can indicate:

[0127] (1) Time range of original features

[0128] DNI monitoring values ​​within 10 minutes prior to the first moment;

[0129] Aerosol forecast data for 30 minutes, including 15 minutes before and 15 minutes after the first moment;

[0130] Solar shortwave radiation monitoring data 15 minutes before the first moment, and solar shortwave radiation forecast data 15 minutes after the first moment, for a total of 30 minutes;

[0131] The first moment is the start time of the forecast interval.

[0132] (2) Lag characteristics of original features

[0133] During the j-th training iteration for the first forecast time point, the DNI monitoring values ​​from 10 minutes prior to the first forecast time point are used as the input data for the model.

[0134] For the j-th round of training for the second forecast time point, the DNI monitoring value 10 minutes before the first forecast time point and the predicted DNI value of the first time point that has already been trained are used as the input data for model training at the second time point;

[0135] or,

[0136] During the j-th round of training for the second prediction time point, the predicted DNI value of the first time point, which has already been trained, is used as the input data for model training at the second time point;

[0137] (3) Derived features corresponding to the original features

[0138] The interaction term for aerosol optical thickness includes humidity;

[0139] It is understandable that by creating an interaction term between aerosol optical thickness and humidity, the degree of influence of the interaction term on DNI prediction can be learned during subsequent training.

[0140] For example, derived features can be created using dot products or cross products.

[0141] (4) Time series features corresponding to the original features

[0142] The rate of change (trend) of DNI values ​​monitored on the ground within 30 minutes prior to the first moment.

[0143] The changing trend of solar shortwave radiation monitoring values ​​provided by the monitoring satellite within a 30-minute interval, 15 minutes before and 15 minutes after the first moment (during model training).

[0144] The changing trends of solar shortwave radiation monitoring values ​​provided by the monitoring satellite 15 minutes before the first moment and the changing trends of solar shortwave radiation forecast values ​​in high-resolution numerical prediction data 15 minutes after the first moment (during model training and model application).

[0145] For example, the rate of change of DNI 10 minutes before the first timestamp (e.g., the DNI prediction value within 6 hours after the first timestamp) can be calculated based on the acquired ground observation data, and the shortwave radiation trend can be determined based on the historical monitoring data of the monitoring satellite. During training, aerosol concentration and shortwave radiation values ​​for 30 minutes, 15 minutes before and after the first timestamp, can be obtained from the acquired historical forecast data.

[0146] It should be noted that the time periods corresponding to each feature indicated in the feature parameters can be learned and updated during the training process, and the feature parameters corresponding to the final training results are used for prediction when the model is applied.

[0147] S205b: Generate model input feature data based on the original feature data, derived feature data, time series feature data, and lagged feature data of the first DNI predicted impact factor.

[0148] The derived features obtained through the feature engineering described above enable the model to capture the synergistic relationships between different key factors during learning. The time series features obtained through the feature engineering described above enable the model to capture the impact of key factors on changes in DNI during learning. The lag features introduced through the feature engineering described above enable the model to capture the short-term dynamic impact of key factors on DNI.

[0149] Based on this scheme, the diversity of DNI prediction influencing factors can be increased through the aforementioned feature engineering. The relationship between various meteorological factors and DNI can be learned from multiple perspectives, so that the subsequently trained model can accurately obtain the DNI prediction results based on various meteorological factors and weight coefficients.

[0150] Optionally, in the method for predicting the DNI of a solar thermal power plant provided in this embodiment of the present disclosure, the above-mentioned S206 may specifically include the following S206a.

[0151] S206a. Based on the model input feature data and the training model, train according to weather type scenarios to obtain the DNI prediction sub-model for each weather type scenario.

[0152] For example, training can be conducted separately for weather scenarios such as sunny days, cloudy days, and dust storms.

[0153] It is understandable that the factors in the DNI prediction influencing factor set may not be exactly the same in different weather scenarios; even when all the factors in the DNI prediction influencing factor set are the same, the weight coefficients of the input features corresponding to each factor are different. For example, in a dust storm scenario, the weight of aerosols is greater than that in a sunny weather scenario.

[0154] Based on this scheme, during the training phase, training can be performed on weather scenario classifications to obtain DNI prediction models under different weather conditions. This makes the trained DNI prediction models more adaptable and improves their prediction performance for different weather scenarios.

[0155] Optionally, in the method for predicting the DNI of a solar thermal power plant provided in this embodiment of the present disclosure, the above-mentioned S206 may specifically include the following S206b1 and S206b2.

[0156] S206b1. Based on the model input data corresponding to the target power station, obtain the model set for training.

[0157] For example, the model set includes: Vector Autoregression (VR), Autoregressive Integrated Moving Average (ARIMA), Long Short-Term Memory (LSTM), Random Forest (RF), Light Gradient Boosting Machine (Light GBM), Extremely Randomized Trees (ERT), Gradient Boosting Decision Tree (GBDT), and Ridge Regression (Ridg).

[0158] S206b2. Train each model in the model set separately, and select the model with the highest prediction performance as the training model for matching the target power station. Use the DNI prediction model trained based on the training model for matching the target power station as the final DNI prediction model for the target power station.

[0159] For example, the training results of multiple models are tested to select the training model that matches the target power station. It is understood that the training models of different power stations in different locations may be the same or different, and this disclosure does not specifically limit this.

[0160] For example, in this embodiment of the disclosure, the Light GBM model is selected as the matching model for the target power plant through performance screening. When training the Light GBM model, the network parameters of the Light GBM model can be optimized through network search. Experiments have verified that the performance is better than EC-HRES forecast and better than the traditional method of relying solely on numerical weather forecast, ground observation or cloud image.

[0161] Based on this approach, multiple types of models can be selected for training during the training process, and the model that best matches each power station can be chosen. When training for different weather scenarios, different types of models can also be selected for training, thereby obtaining sub-models that match each weather scenario, and thus training more accurate sub-models.

[0162] Prediction process:

[0163] Figure 9 A schematic diagram of the processing flow of a method for predicting the Density Incidence (DNI) of a concentrated solar power plant provided in this embodiment of the present disclosure is shown below. Figure 9 As shown, it may include the following S901 to S905:

[0164] S901. Within a preset spatial range centered on the target power station, acquire monitoring data of the predicted impact factor of DNI in the first time period before the current time and forecast data of the predicted impact factor of DNI in the second time period.

[0165] The monitoring data for the DNI prediction impact factors include the monitoring values ​​of DNI and the monitoring values ​​of solar downward shortwave radiation provided by the monitoring satellite; the forecast data for the DNI prediction impact factors include at least one of the following: aerosol optical thickness, PM2.5 concentration and scattering coefficient in aerosol forecast data, geopotential vorticity, cloud cover, temperature, humidity and solar downward shortwave radiation in high-resolution numerical forecast data; the second time period includes the DNI prediction time period of the target power plant and the time period with a preset duration before the prediction time period.

[0166] It should be noted that the relationship between the length of the first time period, the preset duration, and the length of the predicted time period is learned during the model training process.

[0167] For example, the preset duration can be 15 minutes, 30 minutes or 60 minutes, and this embodiment does not specifically limit it.

[0168] For example, monitoring data from 10 minutes prior to the current time can be obtained, forecast data from 30 minutes prior to the current time can be obtained, and forecast data from 6 hours after the current time can be obtained. The forecast data from the 30 minutes prior to the current time can include a single forecast or multiple consecutive forecasts.

[0169] For example, when forecasting the DNI prediction value between 12:00 and 18:00 at 12:00, the value of the first parameter of the forecast between 11:00 and 12:00 can be obtained. The forecast between 11:00 and 12:00 can be a single forecast or multiple forecasts. For example, the first parameter value of the forecast at 11:00 and the first parameter value of the forecast at 11:30 can be obtained, or only the first parameter value of the forecast at 11:30 can be obtained.

[0170] S902. Preprocess the acquired monitoring and forecast data.

[0171] Specifically, after preprocessing, the data is obtained after outlier removal, time alignment, data synchronization, and feature normalization.

[0172] S903. Based on the preprocessed data, construct the model input features for the DNI prediction at the i-th time point within the prediction time period.

[0173] Where i is a positive integer. The model input features include at least one of the following: DNI lag features and DNI predicted impact factors: original features, derived features, and time series features.

[0174] Optionally, the DNI lag feature indicates the DNI value input at the (i-1)th time point when predicting the i-th time point. If i=1, the DNI value includes the DNI monitoring value within the first time period. If i>1, the DNI value includes at least the DNI predicted value before the i-th time point. The original feature indicates the original data of each DNI prediction impact factor. The derived feature includes the interaction term between a DNI prediction impact factor and other DNI prediction impact factors. The time series feature indicates the changing trend of the prediction impact factor.

[0175] For example, time series features include: the changing trend of DNI monitoring values ​​in the first time period, and the changing trend of each DNI prediction influencing factor in the second time period.

[0176] S904. Based on the model input features and the pre-trained DNI prediction model, a sequence of predicted DNI point values ​​is generated sequentially within the prediction time period.

[0177] For example, the predicted time period can be any duration between 0 and 6 hours.

[0178] For example, DNI values ​​can be predicted at 10-minute intervals between 12:00 and 18:00.

[0179] S905. Based on the DNI point prediction value sequence, generate the prediction uncertainty interval and output the final result of the DNI prediction for the target power plant.

[0180] For example, the prediction uncertainty interval of the DNI point prediction value sequence can be generated based on Monte Carlo simulation.

[0181] Based on the DNI point prediction value sequence, the solar thermal power plant can adopt a conservative thermal storage scheme in the low DNI range to avoid excessive power generation and equipment damage; in the high DNI range, the solar thermal power plant can start aggressive scheduling to maximize the solar thermal conversion efficiency.

[0182] This disclosure provides a method for predicting the Density Influence Factor (DNI) of a solar thermal power plant. First, within a predetermined spatial range centered on the target power plant, monitoring data of the DNI prediction influencing factors for a first time period prior to the current time of the target power plant, and forecast data of the DNI prediction influencing factors for a second time period, are acquired. Then, the acquired monitoring and forecast data are preprocessed, and based on the preprocessed data, model input features for predicting the DNI at the i-th time point within the prediction time period are constructed. Next, based on the model input features and a pre-trained DNI prediction model, a sequence of predicted DNI points within the prediction time period is generated sequentially. Finally, based on the DNI point predicted value sequence, a prediction uncertainty interval is generated, and based on the DNI point predicted value sequence and the prediction uncertainty interval, the final result of the DNI prediction for the target power plant is generated and output. Because it deeply integrates aerosol forecast data, shortwave radiation monitoring and forecast values, and multi-source data such as aerosol optical thickness, PM2.5 concentration, and scattering coefficient, it constructs a model input feature that includes at least one of the following: original features, derived features, and time-series features, achieving multi-dimensional synergy and overcoming the limitations of a single data source. By introducing real-time aerosol forecast data and cloud cover, we can simultaneously analyze cloud dynamics and aerosol concentration fluctuations, dynamically quantify the attenuation effect of dust particles on solar radiation, and dynamically capture the fluctuation effect of DNI. By using spatial features and derived features to improve the adaptability of the prediction model to complex weather, we can systematically improve the prediction accuracy of short-term DNI.

[0183] Optionally, in the method for predicting the DNI of a concentrated solar power plant provided in this embodiment of the present disclosure, after S901 above, it may further include the following S906:

[0184] S906. Determine the target weather type corresponding to the target power station at the current time, and call the DNI prediction target sub-model that matches the target weather type.

[0185] The aforementioned S906 may specifically include the following S906a and S906b:

[0186] S906a, Load the DNI prediction sub-model library.

[0187] The DNI prediction sub-model library includes pre-trained DNI prediction sub-models that match various weather patterns.

[0188] S906b. If the current weather pattern is the target weather pattern, select the DNI prediction target sub-model corresponding to the target weather pattern from the DNI prediction sub-model library.

[0189] Optionally, the weather pattern can be determined based on the collected ground monitoring data, such as based on the preprocessed DNI predicted impact factor data, or the weather pattern at the current moment can be determined directly based on the weather forecast. This disclosure does not specifically limit this.

[0190] Therefore, the above-mentioned S904 can be executed through the following S904a:

[0191] S904a: Based on the model input features and the DNI prediction target sub-model, a sequence of predicted DNI point values ​​is generated sequentially within the prediction time period.

[0192] For example, if the current weather type is sunny, then the sunny DNI prediction sub-model is selected; if the current weather type is cloudy, then the cloudy DNI prediction sub-model is selected; if the current weather type is dusty, then the dusty DNI prediction sub-model is selected.

[0193] Based on this approach, for each weather scenario, a scenario-specific training strategy can be used to customize differentiated models for different weather conditions such as sunny, cloudy, and dusty weather, optimizing feature weight allocation. When making predictions, the DNI prediction model that better matches the current weather can be called, thus allowing the selection of prediction parameters based on the weather type, thereby improving the flexibility and adaptability of model predictions. The various pre-trained DNI prediction sub-models can be merged into a DNI prediction sub-model library. When making DNI predictions in different prediction periods, the real-time weather type of the current prediction period can be determined first, and then the sub-model that matches the weather type of the current prediction period can be selected, which can improve the accuracy of real-time short-term DNI predictions and significantly reduce prediction errors under complex weather conditions.

[0194] Example:

[0195] Taking Region 1 as an example, a corresponding DNI prediction model was generated for the solar thermal stations in Region 1. The Light GBM model was ultimately selected for training. The Light GBM model's validation set had a determination coefficient R² of 0.805, demonstrating high training efficiency and accuracy. The hyperparameters of the Light GBM model were optimized using a network search, and the mean absolute error (MAE) and R² were optimized through 5-fold cross-validation, iterating 200 times. The DNI prediction model was trained scenario-specifically using an input sample matrix [number of samples, number of features]. The number of features was 60, including original features, derived features, time-series features, and lagged features. Scenario-specific DNI prediction model training yields DNI prediction sub-models for various weather types, improving DNI prediction performance for specific scenarios.

[0196] After training and generating the DNI prediction model, the training performance was determined based on MAE, Root Mean Square Error (RMSE), and R². The results were obtained by validation on the validation set (January 1, 2024 to April 1, 2024): MAE = 111.11 W / m², RMSE = 186.94 W / m², R² = 0.805, and correlation coefficient = 0.869.

[0197] With scenario-specific training, taking sunny, cloudy, and dusty weather scenarios as examples, the performance of the DNI prediction sub-models for the three scenarios is as follows:

[0198] Sunny day: MAE=95.2 W / m², R²=0.88.

[0199] Cloudy: MAE=125.6 W / m², R²=0.79.

[0200] Dust: MAE=140.3 W / m², R²=0.76.

[0201] Among them, sandstorm weather has a high aerosol variability.

[0202] Compared with the baseline EC-HRES forecast: MAE≈160 W / m², correlation coefficient=0.83.

[0203] Based on the above results, it can be seen that the predictive performance of this application is significantly better than that of the baseline.

[0204] After training and generating the DNI prediction model, uncertainty analysis is performed on each scenario based on Monte Carlo simulation.

[0205] Specifically, normally distributed noise can be added to the input factors of the training model. For example, noise of 0.1 × standard deviation can be added to temperature, humidity, aerosol optical thickness, and satellite radiation. A total of 331,100 simulations were run, and 95% confidence intervals were calculated. The results showed that when the real-time DNI > 800 W / m², the confidence interval of the prediction results was narrower (±30 W / m²), and when the real-time DNI < 200 W / m², the confidence interval of the prediction results was wider (±50 W / m²). That is, a high real-time DNI results in a narrower confidence interval of the prediction results, and a low real-time DNI results in a wider confidence interval of the prediction results.

[0206] The uncertainty results for the three scenarios are as follows:

[0207] Sunny day: High prediction stability, with a 95% confidence interval width ≈ 5% of the predicted value.

[0208] Cloudy: Cloud cover variation rate results in a 95% confidence interval width of approximately 8% of the predicted value.

[0209] Dust storm: Aerosol forecast error makes the 95% confidence interval width approximately 12% of the predicted value.

[0210] Uncertainty intervals can be used for power plant dispatching, providing conservative or aggressive generation plans. The final output is a DNI forecast and uncertainty interval with a resolution of 0-6 hours and 10 minutes.

[0211] Corresponding to the embodiments of the foregoing methods, this disclosure also provides embodiments of the apparatus and the terminal to which it is applied.

[0212] This disclosure discloses an embodiment of a DNI prediction device for a concentrated solar power (CSP) plant, which can be applied to computer equipment, such as servers or terminal devices. The device embodiment can be implemented through software, hardware, or a combination of both. Taking software implementation as an example, as a logical device, it is formed by the DNI prediction processor of the CSP plant loading the corresponding computer program instructions from non-volatile memory into memory for execution. From a hardware perspective, such as... Figure 10 The diagram shown is a hardware structure diagram of a server housing the DNI prediction device for a solar thermal power plant provided in this embodiment of the present disclosure, except... Figure 10 In addition to the processor 1010, memory 1030, network interface 1020, and non-volatile memory 1040 shown, the server or electronic device where the DNI prediction method 1031 of the solar thermal power plant in the embodiment is located may also include other hardware depending on the actual function of the computer device, which will not be described in detail here.

[0213] like Figure 11 As shown, Figure 11This is a block diagram of a device for predicting the direct normal irradiance (DNI) of a solar thermal power plant, provided in an embodiment of this disclosure. The device 1100 includes: a data acquisition module 1101, a data preprocessing module 1102, a model input feature construction module 1103, a DNI prediction module 1104, and an uncertainty analysis module 1105. The data acquisition module 1101 is used to acquire monitoring data of the direct normal irradiance (DNI) prediction influence factor for a first time period prior to the current time and forecast data of the DNI prediction influence factor for a second time period, within a preset spatial range centered on the target power plant. The monitoring data includes the monitored DNI value and the monitored value of downward shortwave solar radiation provided by a monitoring satellite. The forecast data includes at least one of the following: aerosol optical thickness, PM2.5 concentration, and scattering coefficient from aerosol forecast data; and geopotential vorticity, cloud cover, temperature, humidity, and [other data] from high-resolution numerical weather prediction data. The solar shortwave radiation downwards; the second time period includes the predicted time period of the DNI of the target power station and a time period of preset duration before the predicted time period; the data preprocessing module 1102 is used to preprocess the acquired monitoring data and forecast data; the model input feature construction module 1103 is used to construct the model input features for the DNI prediction at the i-th time point within the predicted time period based on the preprocessed data, where i is a positive integer; the model input features include at least one of the following: DNI lag features and DNI prediction influencing factors: original features, derived features, and time series features; the DNI prediction module 1104 is used to generate the DNI point prediction value sequence within the predicted time period based on the model input features and the pre-trained DNI prediction model; the uncertainty analysis module 1105 is used to generate the prediction uncertainty interval based on the DNI point prediction value sequence and output the final result of the DNI prediction of the target power station.

[0214] Optionally, the DNI lag feature indicates the DNI value input at the (i-1)th time point when predicting the i-th time point. If i=1, the DNI value includes the DNI monitoring value within the first time period. If i>1, the DNI value includes at least the DNI predicted value before the i-th time point. The original feature indicates the original data of each DNI prediction impact factor. The derived feature includes the interaction term between a DNI prediction impact factor and other DNI prediction impact factors. The time series feature indicates the changing trend of the prediction impact factor.

[0215] Optionally, the device for DNI prediction in a solar thermal power plant further includes: a sub-model calling module; the sub-model calling module is used to acquire monitoring data of the DNI prediction influencing factors for the first time period and forecast data of the DNI prediction influencing factors for the second time period before the current time, and then load the DNI prediction sub-model library, which includes pre-trained DNI prediction sub-models matching various weather types; if the weather type at the current time is the target weather type, the DNI prediction target sub-model corresponding to the target weather type is selected from the DNI prediction sub-model library; the DNI prediction module 1104 is specifically used to generate a sequence of predicted DNI points within the prediction time period based on the model input features and the DNI prediction target sub-model.

[0216] Optionally, the data preprocessing module is specifically used to perform at least one of the following: removing invalid values ​​from the monitoring data and correcting outliers in the monitoring data; aligning the timestamps of data from different sources for time synchronization; interpolating non-first resolution data to first resolution data, where the first resolution is the preset resolution predicted by DNI; and standardizing the acquired data based on pre-trained normalization parameters.

[0217] Optionally, the device for predicting the DNI of a concentrated solar power (CSP) plant further includes: a DNI prediction influencing factor determination module and a model training module; a data acquisition module, further used to acquire raw data for training the DNI prediction model of the target power plant before sequentially generating the DNI point prediction value sequence within the prediction time period based on the model input features and the pre-trained DNI prediction model, the raw data including at least one of the following: ground monitoring data of the target power plant, reanalysis dataset, aerosol forecast data, shortwave radiation product data provided by satellite, and high-resolution numerical forecast data; ground monitoring data includes DNI monitoring data and meteorological monitoring data; The preprocessing module is used to preprocess the original data to obtain preprocessed training data; the DNI prediction impact factor determination module is used to determine the DNI prediction impact factor based on the preprocessed training data; the model input feature construction module is used to extract the DNI prediction impact factor data at each grid point within a preset spatial range with the target power station as the grid center and a preset size grid from the preprocessed data; and construct model input feature data based on the extracted DNI prediction impact factor data; the model training module is used to train the model based on the model input feature data and the training model to obtain the DNI prediction model.

[0218] Optionally, the DNI prediction impact factor determination module is specifically used to determine the correlation between each factor and DNI based on the preprocessed training data, and select DNI prediction impact factors based on the correlation; with DNI as the target, establish a regression model between each DNI prediction impact factor and DNI, determine the regression coefficient of each DNI prediction impact factor, and select DNI prediction impact factors for training the DNI prediction model based on the regression coefficient of each DNI prediction impact factor.

[0219] The model input feature construction module is specifically used to generate DNI lag features based on the feature parameters and the original features of the extracted first DNI predicted impact factor, and to create derived features and time series features corresponding to the first DNI predicted impact factor. It also generates the model input feature data for this training based on the original feature data, derived feature data, time series feature data, and DNI lag feature data for each DNI predicted impact factor. Here, the first DNI predicted impact factor is any one of the extracted DNI predicted impact factors, and the DNI lag feature indicates the DNI values ​​prior to the i-th time point used to generate the DNI lag feature for the j-th training round at the i-th time point. The numerical value, j, is a positive integer; the derived features include the interaction term of the first DNI predicted impact factor and the second DNI predicted impact factor; the second DNI predicted impact factor is any predicted impact factor that is different from the first DNI predicted impact factor among the extracted DNI predicted impact factors; the time series features indicate the changing trend of the first DNI predicted impact factor; the feature parameters indicate the time range of the original features, the range of DNI values ​​before the i-th time point used to generate the DNI lag feature for the j-th round of training for the i-th time point, the range of the second DNI predicted impact factor included in the derived features, and the time range corresponding to the time series features; the value of the feature parameters is updated after each round of training.

[0220] Optionally, the training module is specifically used to: train the model based on the input feature data and the training model according to the weather type scenario, and obtain the DNI prediction sub-model for each weather type scenario.

[0221] Optionally, the training module is specifically used to: obtain a set of models for training based on the model input data corresponding to the target power station; train each model in the model set separately, select the model with the highest prediction performance as the training model for matching the target power station, and use the DNI prediction model obtained by training the training model for matching the target power station as the final DNI prediction model for the target power station.

[0222] This disclosure provides an apparatus for predicting the Density Influence Factor (DNI) of a solar thermal power plant. First, within a preset spatial range centered on the target power plant, monitoring data of the DNI prediction influence factors for a first time period prior to the current time of the target power plant, and forecast data of the DNI prediction influence factors for a second time period, are acquired. Then, the acquired monitoring and forecast data are preprocessed, and based on the preprocessed data, model input features for predicting the DNI at the i-th time point within the prediction time period are constructed. Next, based on the model input features and a pre-trained DNI prediction model, a sequence of predicted DNI points within the prediction time period is generated sequentially. Finally, based on the DNI point prediction value sequence, a prediction uncertainty interval is generated, and based on the DNI point prediction value sequence and the prediction uncertainty interval, the final result of the DNI prediction for the target power plant is generated and output. Because it deeply integrates aerosol forecast data, shortwave radiation monitoring and forecast values, and multi-source data such as aerosol optical thickness, PM2.5 concentration, and scattering coefficient, it constructs a model input feature that includes at least one of the following: original features, derived features, and time series features, achieving multi-dimensional synergy and overcoming the limitations of a single data source. By introducing real-time aerosol forecast data and cloud cover, we can simultaneously analyze cloud dynamics and aerosol concentration fluctuations, dynamically quantify the attenuation effect of dust particles on solar radiation, and dynamically capture the fluctuation effect of DNI. By using spatial features and derived features to improve the adaptability of the prediction model to complex weather, we can systematically improve the prediction accuracy of short-term DNI.

[0223] Accordingly, this disclosure also provides an electronic device, which includes a processor; a memory; and computer-readable instructions stored on the memory, which, when executed by the processor, implement the DNI prediction method for a solar thermal power plant as described above.

[0224] This disclosure also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method for predicting the DNI of a solar thermal power plant as described above.

[0225] The specific implementation process of the functions and roles of each module in the above device can be found in the implementation process of the corresponding steps in the above method, and will not be repeated here.

[0226] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The modules described as separate components may or may not be physically separate, and the components shown as modules may or may not be physical modules, that is, they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0227] The foregoing has described specific embodiments of this disclosure. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired results. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0228] Other embodiments of this disclosure will readily occur to those skilled in the art upon consideration of the specification and practice of the invention applied herein. This disclosure is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not claimed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this disclosure are indicated by the following claims.

[0229] It should be understood that this disclosure is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this disclosure is limited only by the appended claims.

[0230] The above description is merely a preferred embodiment of this disclosure and is not intended to limit this disclosure. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A method for predicting the Density Incidence (DNI) of a concentrated solar power (CSP) plant, characterized in that, The method includes: Within a predetermined spatial range centered on the target power station, monitoring data of the Direct Normal Irradiance (DNI) prediction impact factor for the first time period prior to the current moment and forecast data of the DNI prediction impact factor for the second time period are acquired. The monitoring data includes the monitored DNI value and the monitored value of solar downward shortwave radiation provided by the monitoring satellite. The forecast data includes at least one of the following: aerosol optical thickness, PM2.5 concentration, and scattering coefficient from aerosol forecast data, and geopotential vorticity, cloud cover, temperature, humidity, and solar downward shortwave radiation from high-resolution numerical weather prediction data. The second time period includes the predicted time period of the DNI of the target power station and a time period of predetermined duration prior to the predicted time period. The acquired monitoring data and forecast data are preprocessed; Based on the preprocessed data, the model input features for predicting DNI at the i-th time point within the prediction period are constructed, where i is a positive integer. The model input features include at least one of the following: DNI lag features and DNI prediction impact factors: original features, derived features, and time series features. The DNI lag features indicate the DNI value input at the (i-1)-th time point when predicting the i-th time point. If i=1, the DNI value includes the DNI monitoring values ​​within the first time period; if i>1, the DNI value includes at least the predicted DNI values ​​before the i-th time point. The original features indicate the original data for each DNI prediction impact factor. The derived features include an interaction term between one DNI prediction impact factor and other DNI prediction impact factors. The time series features indicate the changing trend of the prediction impact factors. Based on the model input features and the pre-trained DNI prediction model, a sequence of predicted DNI points within the prediction time period is generated sequentially. Based on the DNI point prediction value sequence, a prediction uncertainty interval is generated, and the final DNI prediction result of the target power plant is output.

2. The method according to claim 1, characterized in that, After acquiring the monitoring data of the direct normal irradiance (DNI) prediction impact factor for the first time period prior to the current time and the forecast data of the DNI prediction impact factor for the second time period, the method further includes: Load the DNI prediction sub-model library, which includes pre-trained DNI prediction sub-models that match various weather types. If the current weather pattern is the target weather pattern, select the DNI prediction target sub-model corresponding to the target weather pattern from the DNI prediction sub-model library; The DNI prediction model, based on the model input features and a pre-trained model, sequentially generates a sequence of predicted DNI points within the prediction time period, including: Based on the model input features and the DNI prediction target sub-model, a sequence of predicted DNI points within the prediction time period is generated sequentially.

3. The method according to claim 1, characterized in that, The preprocessing of the acquired monitoring data and forecast data includes at least one of the following: Remove invalid values ​​from the monitoring data and correct outliers in the monitoring data; Align timestamps from different data sources for time synchronization; Data that is not at the first resolution is interpolated to data at the first resolution, where the first resolution is the preset resolution predicted by DNI. The acquired data is standardized based on pre-trained normalization parameters.

4. The method according to claim 1, characterized in that, Before generating the sequence of predicted DNI points within the predicted time period based on the model input features and the pre-trained DNI prediction model, the method further includes: Obtain raw data for training the DNI prediction model of the target power plant, wherein the raw data includes at least one of the following: ground monitoring data of the target power plant, reanalysis dataset, aerosol forecast data, shortwave radiation product data provided by satellite, and high-resolution numerical forecast data; wherein the ground monitoring data includes DNI monitoring data and meteorological monitoring data; The original data is preprocessed to obtain preprocessed training data; Based on the preprocessed training data, determine the DNI prediction impact factor; Extract the DNI prediction impact factor data for each grid point within a preset spatial range with the target power plant as the grid center and a preset size grid from the preprocessed data; Based on the extracted DNI prediction impact factor data, the model input feature data is constructed; The DNI prediction model is obtained by training the model based on the input feature data and the training model.

5. The method according to claim 4, characterized in that, The step of determining the DNI prediction impact factor based on the preprocessed training data includes: Based on the preprocessed training data, the correlation between each factor and DNI is determined, and the DNI prediction influencing factor is selected based on the correlation. Using DNI as the target, a regression model is established between each DNI predicted impact factor and DNI, the regression coefficient of each DNI predicted impact factor is determined, and based on the regression coefficient of each DNI predicted impact factor, the DNI predicted impact factor used to train the DNI prediction model is selected.

6. The method according to claim 4, characterized in that, The model input feature data, based on the extracted DNI prediction impact factor data, includes: Based on the feature parameters, according to the original features of the extracted first DNI predicted impact factor, DNI lag features are generated, and derived features and time series features corresponding to the first DNI predicted impact factor are created. Based on the original feature data, derived feature data, time series feature data, and DNI lag feature data of each DNI predicted influencing factor, the model input feature data for this training is generated. Wherein, the first DNI prediction impact factor is any one of the extracted DNI prediction impact factors; the DNI lag feature indicates the DNI value before the i-th time point used to generate the DNI lag feature for the j-th training round at the i-th time point, where j is a positive integer; the derived feature includes the interaction term of the first DNI prediction impact factor and the second DNI prediction impact factor; the second DNI prediction impact factor is any one of the extracted DNI prediction impact factors that is different from the first DNI prediction impact factor; the time series feature includes the changing trend of the first DNI prediction impact factor; the feature parameter indicates the time range of the original feature, the range of the DNI values ​​before the i-th time point used to generate the DNI lag feature for the j-th training round at the i-th time point, the range of the second DNI prediction impact factor included in the derived feature, and the time range corresponding to the time series feature; the value of the feature parameter is updated after each training round.

7. The method according to any one of claims 4 to 6, characterized in that, The process of training the DNI prediction model based on the input feature data and the training model includes: Based on the model input feature data and the training model, the model is trained separately according to weather-type scenarios to obtain a DNI prediction sub-model for each weather-type scenario.

8. The method according to claim 7, characterized in that, The process of training the DNI prediction model based on the input feature data and the training model includes: Based on the model input data corresponding to the target power station, obtain a set of models for training; Each model in the model set is trained separately, and the model with the highest prediction performance is selected as the training model for matching the target power station. The DNI prediction model obtained by training the training model for matching the target power station is used as the final DNI prediction model for the target power station.

9. A device for predicting the DNI (Distributed National Intake) of a concentrated solar power plant, characterized in that, The device includes: a data acquisition module, a data preprocessing module, a model input feature construction module, a direct normal irradiance (DNI) prediction module, and an uncertainty analysis module; The data acquisition module is used to acquire monitoring data of the Direct Normal Irradiance (DNI) prediction impact factor for a first time period prior to the current time and forecast data of the DNI prediction impact factor for a second time period, within a preset spatial range centered on the target power station. The monitoring data includes the monitored DNI value and the monitored value of solar downward shortwave radiation provided by the monitoring satellite. The forecast data includes at least one of the following: aerosol optical thickness, PM2.5 concentration, and scattering coefficient from aerosol forecast data, and geopotential vorticity, cloud cover, temperature, humidity, and solar downward shortwave radiation from high-resolution numerical forecast data. The second time period includes the predicted time period of the DNI of the target power station and a time period of a preset duration prior to the predicted time period. The data preprocessing module is used to preprocess the acquired monitoring data and forecast data; The model input feature construction module is used to construct model input features for the DNI prediction at the i-th time point within the prediction time period based on the preprocessed data, where i is a positive integer. The model input features include at least one of the following: DNI lag features and DNI prediction impact factors: original features, derived features, and time series features. The DNI lag features indicate the DNI value input at the (i-1)-th time point when predicting the i-th time point. If i=1, the DNI value includes the DNI monitoring value within the first time period; if i>1, the DNI value includes at least the DNI prediction value before the i-th time point. The original features indicate the original data for each DNI prediction impact factor. The derived features include an interaction term between a DNI prediction impact factor and other DNI prediction impact factors. The time series features indicate the changing trend of the prediction impact factors. The DNI prediction module is used to generate a sequence of predicted DNI points within the prediction time period based on the model input features and the pre-trained DNI prediction model. The uncertainty analysis module is used to generate a prediction uncertainty interval based on the DNI point prediction value sequence and output the final result of the DNI prediction for the target power plant.

Images