A method for solar thermal DNI prediction based on multi-dimensional forecast bias correction

Abstract

The present invention provides a method for predicting direct normal solar radiation (DNI) for solar thermal power plants based on multidimensional forecast deviation correction, comprising: step 1, performing a cleaning preprocessing operation on DNI time series data using an adaptive threshold method; step 2, fusion analysis of DNI time series data and meteorological factor dimensional expansion based on K-means clustering, and identification of typical days under different weather conditions; step 3, predicting short-term direct normal solar radiation for solar thermal power plants based on a three-dimensional fusion optimization algorithm and error correction. The present invention adopts a multidimensional forecast deviation correction method to preprocess the DNI values ​​of solar thermal power plants that integrate multidimensional meteorological factors, construct a feature matrix by integrating the meteorological factors, perform K-means clustering analysis to obtain typical days under various weather conditions, and finally construct a three-dimensional fusion prediction model to predict DNI, and perform error correction to obtain the final prediction result, thereby improving the accuracy of solar thermal DNI prediction.

Description

Technical Field

[0001] The present invention belongs to the field of photothermal DNI prediction, and in particular relates to a photothermal DNI prediction method based on multi-dimensional forecast deviation correction. Background Art

[0002] In new power systems, the inherent randomness and volatility of wind and photovoltaic power generation pose challenges to system frequency and voltage stability for large-scale grid integration. Compared to traditional wind and photovoltaic technologies, CSP technology offers comparable flexibility and regulation to conventional energy units, enabling the "homogeneous consumption of renewable energy," making it a key technological pillar supporting the integration of high renewable energy penetrations. CSP power stations, with their superior peak-shaving capabilities and energy storage compatibility, hold particularly promising prospects and are poised to become a vital component of the electricity market.

[0003] In solar thermal power forecasting, the most significant factor affecting solar power is direct normal irradiance (DNI). Traditional DNI prediction techniques primarily include numerical weather prediction (NWP) and time series analysis. NWP relies on precise simulations and forward-looking predictions of atmospheric conditions. However, this method often exhibits low forecast accuracy when faced with drastic changes in atmospheric conditions. Time series analysis, on the other hand, relies on historical meteorological and solar irradiance data, using statistical methods to establish a link between historical radiation data and current data to estimate DNI. This method can identify the changing characteristics, trends, and patterns of DNI from time series data, effectively predicting its future trajectory. However, time series analysis focuses primarily on the time series itself and often ignores the influence of external factors. This can lead to significant forecast errors when significant environmental changes occur. Summary of the Invention

[0004] Purpose of the Invention: The present invention aims to address the shortcomings of existing technologies by providing a CSP DNI prediction method based on multi-dimensional forecast deviation correction. This method addresses the low prediction accuracy of existing CSP DNI prediction methods.

[0005] The method of the present invention specifically comprises the following steps:

[0006] Step 1: Perform cleaning and preprocessing of DNI time series data using the adaptive threshold method;

[0007] Step 2: Expand the dimensions of meteorological factors and perform K-means clustering on the DNI time series data to identify typical days under different weather types;

[0008] Step 3: Short-term direct solar normal radiation prediction for CSP plants based on a three-dimensional fusion of a bidirectional gated recurrent unit (BiGRU-Attention) with an attention mechanism, a physics-informed neural network (PINN), a Kolmogorov-Arnold Network (KAN) optimization algorithm, and error correction.

[0009] Step 1 includes:

[0010] Step 1-1: Collect historical DNI time series data and corresponding meteorological data, including temperature, pressure, and cloud cover;

[0011] Step 1-2: Preprocess the historical DNI time series data, calculate the standard score Z-Score, set the threshold according to the distribution of the historical standard score Z-Score, and verify and adjust the threshold iteratively, delete the data above the threshold, and record the preprocessed DNI time series data as ;

[0012] Steps 1-3: DNI time series data Input the modal decomposition algorithm VMD to perform modal decomposition and obtain the DNI decomposition signal , represents the sth sub-signal, and s represents the total number of sub-signals;

[0013] Steps 1-4: Decompose the DNI signal The adaptive adjustment model is input separately to extract the sub-signal features.

[0014] In steps 1-4, the adaptive adjustment model performs the following steps:

[0015] Step 1-4-1, extract extreme values: For each sub-signal imf, calculate and extract the maximum value , minimum value And all extreme points, all extreme points form a set , Represents the nth extreme point extracted, and n represents the total number of extreme points;

[0016] Step 1-4-2, calculate the coefficient of variation:

[0017] Calculate the coefficient of variation based on the mutation events of DNI :

[0018] ,

[0019] in, Indicates the mutation value of DNI, is the average value of historical DNI time series data;

[0020] Step 1-4-3, calculate dynamic window width :

[0021] ,

[0022] in represents the adaptive coefficient;

[0023] Step 1-4-4, screening poles:

[0024] For each extreme point, calculate the Euclidean distance between the extreme point and the nearest neighbor extreme point. If the Euclidean distance is greater than the dynamic window width, ,satisfy , then the extreme points are retained and counted as valid extreme points ,in Indicates the kth extreme point and the k+1th extreme point Otherwise, the extreme points are discarded; finally, all the extreme points that meet the conditions are retained to obtain the effective extreme point set , represents the fth effective extreme point;

[0025] Steps 1-4-5, feature enhancement, include the following steps:

[0026] Steps 1-4-5-1 use the moving weighted average method and Holt's linear trend method to calculate the moving average or exponential smoothing of the DNI time series data:

[0027] The formula for the moving weighted average method is:

[0028] ,

[0029] The formula for Holt's linear trend method is:

[0030] ,

[0031] in represents the moving weighted average at time t, used to predict future values, Represents weight, which is used to assign different weights to observations at different time points. represents the observed value of DNI at time i, which is used to calculate the weighted average value, represents the smoothed level value at time t, It represents the horizontal smoothing coefficient, which is usually set to (0,1). In this invention, it is set to 0.5. is the smoothing level value at time t-1, is the smoothed trend value at time t-1; represents the observed value of DNI at time t.

[0032] Steps 1-4-5-2, use Fourier transform to extract seasonal components and add meteorological data as auxiliary features;

[0033] Use the Discrete Fourier Transform DFT:

[0034] ,

[0035] in, is the Fourier coefficient of the kth frequency component, is the nth data point in the time series, N is the length of the time series, and k is the frequency index, ranging from 0 to N-1;

[0036] From the Fourier transform results, select the low-frequency components representing seasonal changes, and use the inverse Fourier transform IFFT to transform the selected seasonal components back to the time domain:

[0037] ;

[0038] Steps 1-4-6, handling outliers: Use the interquartile range (IQR) and Z-score method to detect and remove outliers:

[0039] ,

[0040] ,

[0041] in, and Indicates an outlier, represents the average DNI value, represents the DNI standard deviation; represents the median DNI value, The interquartile range of DNI values ​​refers to the 50% range from the first quartile to the middle of the third quartile;

[0042] Steps 1-4-7, multi-scale analysis: Use wavelet transform to perform multi-scale decomposition of DNI signals:

[0043] ,

[0044] in Represents the wavelet transform coefficient, which refers to the characteristics of the signal at scale a and position b. is the complex conjugate of the wavelet function, a is the scale parameter, which controls the expansion and contraction of the wavelet function, and b is the translation parameter, which controls the position of the wavelet function on the time axis. represents the function of t with respect to the change of DNI; d is the integral symbol.

[0045] The present invention also provides an electronic device, comprising a processor and a memory, wherein the memory stores program code, and when the program code is executed by the processor, the processor executes the steps of the method.

[0046] The present invention also provides a storage medium storing a computer program or instruction, which executes the steps of the method when the computer program or instruction is run on a computer.

[0047] The present invention inputs historical meteorological factor data into an adaptive selection model for deviation correction and data preprocessing, and constructs a feature matrix. The typical weather day is obtained through the K-means clustering method as the input data of the solar thermal DNI prediction model. A short-term direct normal solar radiation prediction model for solar thermal power stations based on the three-dimensional fusion BiGRU-Attention+PINN+KAN optimization algorithm and error correction is designed, which can effectively improve the accuracy of solar thermal DNI prediction under different weather conditions.

[0048] The present invention has the following beneficial effects: First, it optimizes the preprocessing step for data cleaning. This method removes outliers by designing an adaptive threshold model, introduces other meteorological factors into conventional DNI sequences through feature enhancement, and performs multi-scale analysis, making the preprocessed data more distinctive and operational.

[0049] Second, the clustering method is optimized. The present invention designs a feature matrix to perform K-means clustering operations, introduces multidimensional factors, and realizes the extraction of typical weather days, thus playing the role of multidimensional forecasting.

[0050] Third, the prediction model uses three dimensions to further improve accuracy. This paper designs a three-dimensional prediction model that significantly improves the prediction accuracy compared to traditional single models. It uses the BiGRU-Attention, PINN, and KAN methods for three-dimensional prediction and performs error correction on the results to obtain more accurate prediction results. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 It is a technical route flow chart of the three steps of the method of the present invention.

[0052] Figure 2 This is the rainy weather clustering heat map obtained by K-means clustering in the method of the present invention.

[0053] Figure 3 This is the cloudy weather clustering heat map obtained by K-means clustering in the method of the present invention.

[0054] Figure 4 This is the sunny weather clustering heat map obtained by K-means clustering in the method of the present invention.

[0055] Figure 5 Schematic diagram of the GRU structure of the method of the present invention.

[0056] Figure 6 Schematic diagram of the BiGRU structure of the method of the present invention.

[0057] Figure 7 This is a parameter setting diagram for the model training algorithm of the present invention.

[0058] Figure 8 It is a schematic diagram of the prediction results for a typical sunny day in the present invention.

[0059] Figure 9 It is a schematic diagram of the prediction results for a typical cloudy day in the present invention.

[0060] Figure 10 It is a schematic diagram of the prediction results for a typical day with complex weather in the present invention.

[0061] Figure 11 Schematic diagram of parameter evaluation of prediction results of the present invention. DETAILED DESCRIPTION

[0062] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments, and the above and / or other advantages of the present invention will become more apparent.

[0063] The embodiment of the present invention provides a method for predicting CSP DNI based on multi-dimensional forecast deviation correction, comprising the following steps:

[0064] like Figure 1 Shown is a technical route flow chart of the three main steps of the present invention (NWP refers to Numerical Weather Prediction).

[0065] Step 1: Perform cleaning and preprocessing of DNI time series data using the adaptive threshold method;

[0066] Step 1-1: Collect historical DNI time series data and corresponding meteorological data, such as temperature, air pressure, cloud cover, etc.

[0067] Step 1-2: Preprocess the historical DNI time series data, calculate the Z-Score, and find the Z-Score value corresponding to the selected percentile based on the distribution of the historical Z-Score. The dynamic threshold is selected as 95%. The threshold is verified and adjusted iteratively. The preprocessed DNI time series data is recorded as ;

[0068] Steps 1-3: DNI time series data Input VMD modal decomposition algorithm to perform modal decomposition and obtain DNI decomposition signal ;

[0069] Steps 1-4: Decompose the DNI signal Separately input the adaptive adjustment model to extract sub-signal features, reduce the interference of small fluctuations contained in the sub-signals, and enhance the robustness of the system;

[0070] Among them, one of the main innovations of the present invention is the pole adaptive selection model in steps 1-4, which is used to process the VMD model generated The model can filter out the interference of small fluctuations contained in the sub-signals and integrate multiple factors such as season and weather to perform multi-dimensional and multi-scale analysis of DNI. Its operation process is as follows:

[0071] Step 1-4-1, extract extreme values: For each sub-signal imf, calculate and extract the maximum value , minimum value And all extreme points, all extreme points form a set , Represents the nth extreme point extracted, and n represents the total number of extreme points;

[0072] Step 1-4-2, calculate the coefficient of variation:

[0073] Calculate the coefficient of variation based on the mutation events of DNI (Variance Coefficient):

[0074] ,

[0075] in, Indicates the mutation value of DNI, which is generally approximately equal to its standard deviation , is the average value of historical DNI time series data.

[0076] Step 1-4-3, calculate the dynamic window width:

[0077] The window width is dynamically adjusted according to the coefficient of variation of the DNI time series data. When the data volatility is large, the window width should be increased accordingly to reduce misjudgment; when the volatility is small, the window width should be reduced to more accurately identify the extreme points. An adaptive coefficient is defined ,Since the data fluctuation is small in each step, it is necessary to identify the data features more accurately. Choosing a smaller window width can capture the detailed features in the data. The range is set to According to the coefficient of variation and adaptive coefficient , calculate the dynamic window width .

[0078] Dynamic window width The calculation formula is:

[0079] ;

[0080] Step 1-4-4, screening extreme points: For each extreme point, calculate the Euclidean distance between the extreme point and the nearest neighbor extreme point. If the Euclidean distance is greater than the dynamic window width ,satisfy , then the extreme points are retained and counted as valid extreme points ,in Indicates the kth extreme point and the k+1th extreme point Otherwise, the extreme points are discarded; finally, all the extreme points that meet the conditions are retained to obtain the effective extreme point set , represents the fth effective extreme point;

[0081] Steps 1-4-5, feature enhancement: By introducing additional relevant features, the complexity of DNI changes can be better captured. Specifically, the following steps are included:

[0082] In steps 1-4-5-1, the moving average or exponential smoothing of the DNI time series data is calculated to capture the long-term trend, using the moving weighted average method and the Holt linear trend method respectively.

[0083] The formula for the moving weighted average method is:

[0084] ,

[0085] The formula for Holt's linear trend method is:

[0086] ,

[0087] in represents the moving weighted average at time t, used to predict future values, Represents weight, which is used to assign different weights to observations at different time points. represents the observed value of DNI at time i, which is used to calculate the weighted average value, represents the smoothed level value at time t, It represents the horizontal smoothing coefficient, which is usually set to (0,1). In this invention, it is set to 0.5. is the smoothing level value at time t-1, is the smoothed trend value at time t-1;

[0088] Steps 1-4-5-2, use Fourier transform to extract seasonal components, and add meteorological data such as cloud cover, humidity, and temperature as auxiliary features;

[0089] Use the Discrete Fourier Transform DFT:

[0090] ,

[0091] in, is the Fourier coefficient of the kth frequency component, is the nth data point in the time series, N is the length of the time series, and k is the frequency index, ranging from 0 to N-1;

[0092] From the Fourier transform results, low-frequency components representing seasonal changes are selected. Since the subsequent DNI prediction is on an ultra-short-term scale, the seasonal threshold is set to 1 / 12, that is, the daily collection step is 2 hours. Anything below this threshold is due to daily seasonal factors, otherwise it is due to the influence of other meteorological factors.

[0093] The selected seasonal components are converted back to the time domain using an inverse Fourier transform (IFFT):

[0094] ;

[0095] Steps 1-4-6, handling outliers: Use IQR (interquartile range) and Z-score methods to detect and replace outliers. Removing outliers can reduce their impact on extreme point identification and prediction model training:

[0096] ,

[0097] ;

[0098] Steps 1-4-7, multi-scale analysis: Use wavelet transform to perform multi-scale decomposition of the DNI signal. Signals at different time scales may contain different information, and multi-scale analysis helps identify this information.

[0099] Use wavelet transform to perform multi-scale decomposition of DNI signals:

[0100] ;

[0101] Steps 1-4-8, ensemble learning: Combining multiple prediction models, such as random forest and gradient boosting, ensemble learning can reduce the bias and variance of a single model and improve the accuracy of prediction by combining the prediction results of multiple models.

[0102] Random Forest: ;

[0103] Gradient Boosting: ;

[0104] Through the combined application of the aforementioned techniques and methods, the performance of the adaptive extreme point selection model in DNI prediction can be significantly improved. By implementing this invention, an algorithm can be obtained that accurately identifies extreme, sudden meteorological environments within historical DNI time series data. By processing the sub-signals output by the VMD model using a multi-scale adaptive extreme point model that incorporates multidimensional factors, the interference from small fluctuations within the sub-signals is significantly reduced. This enables automated identification of sudden meteorological environments based on historical DNI data, significantly improving the accuracy of complete identification of sudden meteorological environments.

[0105] Step 2: Fusion analysis of DNI time series data based on K-means clustering and meteorological factor dimension expansion and identification of typical days with different weather conditions;

[0106] Step 2-1: Expand the dimensions and perform weighted analysis on meteorological factors, and construct a feature matrix based on derived features.

[0107] Among them, one of the main innovations of the present invention is the dimensional expansion of meteorological factors in step 2-1, as well as the weighted analysis and construction of a feature matrix. The operation process is as follows:

[0108] Step 2-1-1: Expand the dimensions of meteorological factors, perform weighted analysis, and construct a feature matrix;

[0109] Step 2-1-1: Collect basic meteorological data including cloud cover, pressure, temperature, humidity, and wind speed.

[0110] Step 2-1-2: Preprocess the data, use interpolation and mean filling to handle missing values, convert data of different dimensions to the same scale, and use Min-Max normalization:

[0111] ;

[0112] Step 2-1-3, calculate the derived characteristics and use the difference method to calculate the change rate of cloud cover, air pressure, temperature, humidity, and wind speed respectively ,in, Indicates the change value, Indicates the value corresponding to the next time point. Divide by the time interval Get the rate of change of meteorological data;

[0113] Step 2-1-4: Use Pearson correlation analysis to determine the influence coefficient of various meteorological data on CSP DNI, and perform weighted analysis:

[0114] ,

[0115] Where r represents the correlation coefficient, Represents the values ​​of different meteorological data (wind speed, air pressure, temperature, cloud cover, humidity) at time i, represents the average value of the entire set of data. Indicates DNI data at time i, It represents the average value of DNI data of the whole set of data, from which the influence coefficient of different meteorological factors on DNI can be calculated;

[0116] Step 2-1-5, arrange the processed features by columns to form a matrix X: ;

[0117] in Refers to the original values ​​of different meteorological factors within a certain step length (the step length in this invention is 5 minutes), Refers to the rate of change of meteorological data, Refers to the correlation coefficient of light, heat, and nitrogen dioxide (DNI) in the Pearson analysis of meteorological data, specifically expressed as:

[0118] ,

[0119] in, Refers to the data of wind speed, air pressure, temperature, cloud cover and humidity at time i respectively;

[0120] ,

[0121] in 、 、 、 、 They respectively refer to the rate of change of wind speed, air pressure, temperature, cloud cover and humidity at time point i;

[0122] ,

[0123] in 、 、 、 、 They respectively refer to the correlation coefficients between wind speed and DNI, pressure and DNI, temperature and DNI, cloud cover and DNI, and humidity and DNI;

[0124] Through the above steps, the original meteorological data can be converted into an information-rich multidimensional feature matrix, providing effective input for the subsequent K-means clustering analysis.

[0125] Step 2-2, take matrix X as input and perform K-means clustering to identify typical days under different weather types. First, we get clusters of several typical weather types, as follows: Figure 2 、 Figure 3 、 Figure 4 As shown, typical days under different weather types were selected, namely, July 25th with sunny weather, October 26th with cloudy weather, and December 24th with complex weather. The specific steps include the following:

[0126] Step 2-2-1, perform K-means clustering;

[0127] This embodiment uses K-means clustering to identify typical days under different weather types.

[0128] The goal of using the K-means algorithm is to minimize the following objective function:

[0129] ,

[0130] Among them, K represents the number of clusters, is the value of the Within-Cluster Sum of Square (WCSS), It is clusters, It is The cluster centers of the clusters, is the jth data point and cluster centers The square of the Euclidean distance between them;

[0131] K-means clustering includes the following steps:

[0132] Step 2-2-1-1, Initialization: Randomly select Q data points as the initial cluster centers:

[0133] ;

[0134] in Indicates the The initial cluster centers of the clusters, Indicates the The first data point in a cluster;

[0135] Step 2-2-1-2, assignment: For each data point, calculate the distance between the data point and each cluster center, and assign the data point to the cluster represented by the nearest cluster center:

[0136] ;

[0137] in, Represents the jth data point With the The cluster centers The Euclidean distance between Represents the jth data point and the cluster center of the qth cluster The Euclidean distance between

[0138] Step 2-2-1-3, recalculate the cluster center of each cluster;

[0139] Step 2-2-1-4, iteration: Repeat steps 2-2-1-2 to 2-2-1-3 until the condition is met, which means that the cluster center no longer changes significantly, that is, the fluctuation is within the threshold of 0.01, or the maximum number of iterations is reached, which is set to 100;

[0140] Step 2-2-1-5, analyze the data points in each cluster, determine the weather type represented by each cluster, and in each cluster, select the data point closest to the cluster center as the typical day.

[0141] In step 2-2-1-3, the cluster center of each cluster is recalculated using the following formula: .

[0142] Step 3: Short-term direct solar normal radiation prediction for CSP plants based on the three-dimensional fusion BiGRU-Attention+PINN+KAN optimization algorithm and error correction;

[0143] Step 3-1: Use the gated recurrent unit (GRU) as the underlying core network and optimize it to a bidirectional gated recurrent unit (BiGRU-Attention) with an attention mechanism.

[0144] like Figure 5 The figure shows the GRU model structure diagram designed by the present invention, wherein: and They are update gate and reset gate respectively; is the hidden state at time t; is the hidden state at time t-1; is the sigmoid activation function; is the input of the gated recurrent unit GRU at time t; is the hyperbolic tangent function. The forward propagation expression is obtained from the structure diagram of the gated recurrent unit GRU:

[0145] ,

[0146] ,

[0147] ,

[0148] ,

[0149] ,

[0150] in, is the weight parameter; is the output of the GRU network at time t; is the bias parameter; is the candidate hidden state; Multiply matrices element-wise; is the output layer weight;

[0151] like Figure 6 The figure shows the BiGRU model structure diagram designed by the present invention. The information transmission of the bidirectional gated recurrent unit BiGRU is divided into two directions: forward and backward. The output vector transformed from the forward input data is combined with the output vector transformed from the backward input data to obtain the final output vector. The hidden layer state of the bidirectional gated recurrent unit BiGRU is The expression is:

[0152] ,

[0153] ,

[0154] ,

[0155] in, It is the operation process of the GRU network; , Represent the weight and state of the forward hidden layer at time t respectively; , Respectively represent the backward hidden layer weight and state at time t; is the bias of the hidden layer state at time t;

[0156] When introducing the attention mechanism, the output of the bidirectional gated recurrent unit BiGRU is set to ,in It is the concatenation of the forward output and backward output of the bidirectional gated recurrent unit BiGRU at the tth time step;

[0157] First calculate the attention weight for each time step:

[0158] ,

[0159] ,

[0160] in represents the score function added to the attention model, and To learn the parameters, is a weight variable, T represents matrix transpose, represents the attention weight at time step t;

[0161] Calculate the weighted output c:

[0162] ;

[0163] This result c is used as the input of the subsequent network for further operation.

[0164] In step 3-2, the output c of step 3-1 is used as input and input into the physical information neural network PINN for optimization.

[0165] In step 3-3, the output result of step 3-2 is used as input and input into the knowledge enhancement network KAN for further optimization to obtain a preliminary prediction value.

[0166] In step 3-4, the error sequence is obtained through the preliminary prediction value, and the error sequence is predicted using the bidirectional gated recurrent unit BiGRU. The predicted error sequence is arithmetic summed with the preliminary prediction result to obtain the final prediction result.

[0167] One of the main innovations of this invention is the construction of a three-dimensional fusion BiGRU-Attention+PINN+KAN optimization algorithm. This algorithm uses the traditional BiGRU bidirectional gated recurrent unit as the underlying network, sequentially inputting the physical information neural network PINN and the knowledge augmentation network KAN, further improving the prediction results. The specific operations of each module are as follows:

[0168] Physics-Informed Neural Networks (PINNs) are machine learning models that combine deep learning and physics knowledge. Unlike traditional data-driven neural networks, PINNs use physical laws to guide the model during learning, thereby improving its generalization capabilities. The main implementation steps of step 3-2 are as follows:

[0169] Step 3-2-1, build a neural network framework;

[0170] Building a neural network that includes physical information ,in are the parameters of the network, is the predicted physical quantity, and They are spatial variables and time variables; neural network Able to receive the weighted output c from step 3-1 and incorporate geographical location factors and cloud cover, temperature and meteorological data;

[0171] Step 3-2-2, integrating the laws of physics;

[0172] Integrate the law of conservation of energy, the radiation transfer equation and the atmospheric radiation model and express it in the form of partial differential equations, where the law of conservation of energy is expressed as: ;

[0173] in represents the rate of change of energy density, It is the divergence of energy flow, which represents the net amount of energy flowing into or out of a volume element.

[0174] The radiation transfer equation describes the propagation of radiation in a medium, including absorption, emission, and scattering processes, and is expressed as:

[0175] ;

[0176] in, represents the radiation intensity, represents the starting angle, is the radiation angle, s is the distance along the propagation direction, is the absorption coefficient, B is the blackbody radiation intensity, is the scattering phase function, which indicates the scattering angle The scattered light intensity in the direction, is the solid angle element; I is the irradiation;

[0177] The atmospheric radiation model is used to describe the transmission process of radiation in the atmosphere and is expressed as:

[0178] ,

[0179] Among them, z represents the vertical height, represents the absorption coefficient, represents the scattering coefficient, represents the scattering phase function, represents the scattering angle, represents the scattering intensity, represents the atmospheric emission intensity at height z;

[0180] Step 3-2-3, define the loss function;

[0181] Assume that the partial differential equations for the radiation transfer equation and the atmospheric radiation model in step 3-2-2 are expressed as:

[0182] ,

[0183] where L is an operator involving derivative operations.

[0184] The residual of the partial differential equation in step 3-2-2 is used as part of the loss function, and the data loss, physical damage and regularization terms are defined separately; the loss function L is a weighted combination of data loss and partial differential equation (PDE) residual loss.

[0185] Assume the original GRU loss function is the mean square error ;

[0186] PDE residual loss Expressed as:

[0187] ,

[0188] ,

[0189] in, It is the weight coefficient for balancing data loss and PDE residual loss. Because it requires physical constraints, the value of this embodiment is 0.6.

[0190] Step 3-2-5, training the network;

[0191] Use the loss function with physical information to train the neural network, and optimize the data loss and physical loss to obtain the optimized prediction results.

[0192] First, perform forward propagation and calculate the network output and PDE residuals ;

[0193] Then, the loss function is obtained by weighted combination ;

[0194] Finally, perform back propagation to calculate the network parameters The gradient of , and update the network parameters according to the gradient descent method , the expression is:

[0195] ,

[0196] in, represents the learning rate, Indicates network parameters The gradient, It means using the formula on the right to update the network parameters to obtain the preliminary predicted DNI sequence.

[0197] In step 3-3, the Knowledge-Aware Network (KAN) is a knowledge-enhanced neural network architecture that optimizes the model's training results by integrating prior knowledge. Its main implementation steps are as follows:

[0198] Step 3-3-1, extract knowledge;

[0199] Relevant knowledge that affects the prediction of DNI time series data, such as weather conditions, seasonal changes, and diurnal variations of solar radiation, is extracted from existing empirical formulas or prior knowledge, and the relevant knowledge is encoded into an operational mathematical form, such as the form of prior probability distribution and secondary principal loss function.

[0200] Step 3-3-2, knowledge embedding;

[0201] Assume the original GRU loss function is the mean square error , the loss function after integrating knowledge is modified as follows:

[0202] ,

[0203] The residual of the partial differential equation in step 3-2-2 is used as part of the loss function, where is the loss term related to knowledge, is the weight coefficient that balances data loss and knowledge loss, Defined as:

[0204] ;

[0205] in, is the output of the model being trained at the i-th sample point, is the output of the prior knowledge model in the GRU network at time i;

[0206] Prediction results for:

[0207] ,

[0208] in, is the output of the GRU network, is to adjust the weight of knowledge influence, thereby obtaining further prediction results.

[0209] In terms of parameter setting, the present invention divides a total of 16,000 groups of valid experimental data into training sets and test sets, with a division ratio of 80% for training and 20% for testing. The input features of the present invention are based on a 5-minute time step, and the direct normal radiation (DNI) and meteorological data in the first 2 hours are used to predict the DNI value for the next 5 minutes. During the model training process, the Adam optimization algorithm is used to update the network weights. The hyperparameters of the model are optimized by the grid search method, and it is finally determined that the network structure contains a hidden layer, which contains 50 neurons. It can be seen from the change of the loss function and the learning rate that both the training and verification losses show a downward trend and eventually tend to be stable, and the curves of the two are almost completely consistent. This shows that under the current parameter configuration, the model shows good performance after 50 iterations. Therefore, all data (including the test set) are used for model training, and the training is terminated after 50 iterations to obtain a final training model with a training round of 50. The adjusted model parameter settings are as follows Figure 7 shown.

[0210] like Figure 8 、 Figure 9 、 Figure 10 The following figures show the three-dimensional model predictions for typical sunny, cloudy, and mixed weather days, as well as the predictions after error correction, compared with the actual DNI values. As can be seen from the figures, while certain systematic errors exist when using only the three-dimensional model for prediction, error correction significantly reduces these errors, improving prediction accuracy. Figure 11 For parameter evaluation after model training,

[0211] The present invention provides a CSP DNI prediction method based on multidimensional forecast deviation correction. There are numerous methods and approaches for implementing this technical solution. The above is merely a preferred embodiment of the present invention. It should be noted that those skilled in the art may make various improvements and modifications without departing from the principles of the present invention, and such improvements and modifications are also within the scope of protection of the present invention. Any components not specified in this embodiment may be implemented using existing technologies.

Claims

1. A method for solar thermal DNI prediction based on multi-dimensional forecast deviation correction, characterized in that: The following steps are involved: Step 1: Perform cleaning and preprocessing of DNI time series data using the adaptive threshold method; Step 2: Expand the dimensions of meteorological factors and perform K-means clustering on the DNI time series data to identify typical days under different weather types; Step 3: Short-term direct solar normal radiation prediction for CSP plants based on a three-dimensional fusion of a bidirectional gated recurrent unit (BiGRU-Attention) with an attention mechanism, a physical information neural network (PINN), a Kolmogov-Arnold network (KAN) optimization algorithm, and error correction. Step 1 includes: Step 1-1: Collect historical DNI time series data and corresponding meteorological data, including temperature, pressure, and cloud cover; Step 1-2: Preprocess the historical DNI time series data, calculate the standard score Z-Score, set the threshold according to the distribution of the historical standard score Z-Score, and perform threshold verification and adjustment through iteration, delete the data above the threshold, and record the preprocessed DNI time series data as ; Steps 1-3: DNI time series data Input the modal decomposition algorithm VMD to perform modal decomposition and obtain the DNI decomposition signal , represents the sth sub-signal, and s represents the total number of sub-signals; Steps 1-4: Decompose the DNI signal The adaptive adjustment model is input separately to extract the sub-signal features.

2. The method according to claim 1, characterized in that In steps 1-4, the adaptive adjustment model performs the following steps: Step 1-4-1, extract extreme values: For each sub-signal imf, calculate and extract the maximum value , minimum value And all extreme points, all extreme points form a set , Represents the nth extreme point extracted, and n represents the total number of extreme points; Step 1-4-2, calculate the coefficient of variation: Calculate the coefficient of variation based on the mutation events of DNI : ， in, Indicates the mutation value of DNI, is the average value of historical DNI time series data; Step 1-4-3, calculate dynamic window width : ， in represents the adaptive coefficient; Step 1-4-4, screening poles: For each extreme point, calculate the Euclidean distance between the extreme point and the nearest neighbor extreme point. If the Euclidean distance is greater than the dynamic window width, ,satisfy , then the extreme points are retained and counted as valid extreme points ,in Indicates the kth extreme point and the k+1th extreme point Otherwise, the extreme points are discarded; finally, all the extreme points that meet the conditions are retained to obtain the effective extreme point set , represents the fth effective extreme point; Steps 1-4-5, feature enhancement, include the following steps: Steps 1-4-5-1 use the moving weighted average method and Holt's linear trend method to calculate the moving average or exponential smoothing of the DNI time series data: The formula for the moving weighted average method is: ， The formula for Holt's linear trend method is: ， in represents the moving weighted average at time t, used to predict future values, Represents weight, which is used to assign different weights to observations at different time points. represents the observed value of DNI at time i, which is used to calculate the weighted average value, represents the smoothed level value at time t, represents the horizontal smoothing coefficient, is the smoothing level value at time t-1, is the smoothed trend value at time t-1; represents the observed value of DNI at time t; Steps 1-4-5-2, use Fourier transform to extract seasonal components and add meteorological data as auxiliary features; Use the Discrete Fourier Transform DFT: ， in, is the Fourier coefficient of the kth frequency component, is the nth data point in the time series, N is the length of the time series, and k is the frequency index, ranging from 0 to N-1; From the Fourier transform results, select the low-frequency components representing seasonal changes, and use the inverse Fourier transform IFFT to transform the selected seasonal components back to the time domain: ； Steps 1-4-6, dealing with outliers: Use the interquartile range (IQR) and Z-score method to detect and remove outliers: ， ， in, and Indicates an outlier, represents the average DNI value, represents the DNI standard deviation; represents the median DNI value, It represents the interquartile range of DNI values; Steps 1-4-7, multi-scale analysis: Use wavelet transform to perform multi-scale decomposition of DNI signals: ， in Represents the wavelet transform coefficient, which refers to the characteristics of the signal at scale a and position b. is the complex conjugate of the wavelet function, a is the scale parameter that controls the expansion and contraction of the wavelet function, and b is the translation parameter. represents the function of t with respect to the change of DNI; d is the integral symbol.

3. The method according to claim 2, characterized in that Step 2 includes: Step 2-1: Expand the dimensions and perform weighted analysis on meteorological factors, and construct a feature matrix based on derived features. This includes the following steps: Step 2-1-1, collect meteorological data including cloud cover, air pressure, temperature, humidity and wind speed; Step 2-1-2: Preprocess the data, use interpolation and mean filling to handle missing values, convert data of different dimensions to the same scale, and use Min-Max normalization: ， in, represents the normalized value, Represents the data in row m and column n of the original data; Step 2-1-3, calculate the derived characteristics and use the difference method to calculate the change rate of cloud cover, air pressure, temperature, humidity, and wind speed respectively ,in, Indicates the change value, Indicates the value corresponding to the next time point. Divide by the time interval Get the rate of change of meteorological data; Step 2-1-4: Use Pearson correlation analysis to determine the influence coefficient of various meteorological data on CSP DNI, and perform weighted analysis: ， Where r represents the correlation coefficient, Represents the values ​​of different meteorological data at time i, represents the average value of the entire set of data. Indicates DNI data at time i, It represents the average value of DNI data for the entire set of data; Step 2-1-5, arrange the processed features in columns to form a matrix X: ， in Refers to the original values ​​of different meteorological factors within the step length, Refers to the rate of change of meteorological data, Refers to the correlation coefficient of light, heat, and nitrogen dioxide (DNI) in the Pearson analysis of meteorological data, specifically expressed as: ， in, Refers to the data of wind speed, air pressure, temperature, cloud cover and humidity at time i respectively; ， in 、 、 、 、 They respectively refer to the rate of change of wind speed, air pressure, temperature, cloud cover and humidity at time point i; ， in 、 、 、 、 They respectively refer to the correlation coefficients between wind speed and DNI, pressure and DNI, temperature and DNI, cloud cover and DNI, and humidity and DNI; Step 2-2 uses the matrix X as input and performs K-means clustering to identify typical days under different weather types. This includes the following steps: Step 2-2-1, perform K-means clustering: The goal of using the K-means algorithm is to minimize the following objective function: ， Among them, K represents the number of clusters, is the value of the within-cluster sum of squares, It is clusters, It is The cluster centers of the clusters, is the jth data point and cluster centers The square of the Euclidean distance between them; K-means clustering includes the following steps: Step 2-2-1-1, Initialization: Randomly select Q data points as the initial cluster centers: ； in Indicates the The initial cluster centers of the clusters, Indicates the The first data point in a cluster; Step 2-2-1-2, assignment: For each data point, calculate the distance between the data point and each cluster center, and assign the data point to the cluster represented by the nearest cluster center: ； in, Represents the jth data point With the The cluster centers The Euclidean distance between Represents the jth data point and the cluster center of the qth cluster The Euclidean distance between Step 2-2-1-3, recalculate the cluster center of each cluster; Step 2-2-1-4, iteration: repeat steps 2-2-1-2 to 2-2-1-3 until the conditions are met; Step 2-2-1-5, analyze the data points in each cluster, determine the weather type represented by each cluster, and in each cluster, select the data point closest to the cluster center as the typical day.

4. The method according to claim 3, characterized in that In step 2-2-1-3, the cluster center of each cluster is recalculated using the following formula: 。 5. The method according to claim 4, characterized in that Step 3 includes: Step 3-1: Use the gated recurrent unit (GRU) as the underlying core network and optimize it to a bidirectional gated recurrent unit (BiGRU-Attention) with an attention mechanism. and They are update gate and reset gate respectively; is the hidden state at time t; is the hidden state at time t-1; is the sigmoid activation function; is the input of the gated recurrent unit GRU at time t; is the hyperbolic tangent function, and the forward propagation expression is obtained from the structure diagram of the gated recurrent unit GRU: ， ， ， ， ， in, is the weight parameter; is the output of the GRU network at time t; is the bias parameter; is the candidate hidden state; Multiply matrices element-wise; is the output layer weight; The information transmission of the bidirectional gated recurrent unit BiGRU is divided into two directions: forward and backward. The output vector transformed from the forward input data is combined with the output vector transformed from the backward input data to obtain the final output vector. The hidden layer state of the bidirectional gated recurrent unit BiGRU is The expression is: ， ， ， in, It is the operation process of the GRU network; , Represent the weight and state of the forward hidden layer at time t respectively; , Respectively represent the backward hidden layer weight and state at time t; is the bias of the hidden layer state at time t; When introducing the attention mechanism, the output of the bidirectional gated recurrent unit BiGRU is set to ,in It is the concatenation of the forward output and backward output of the bidirectional gated recurrent unit BiGRU at the tth time step; First calculate the attention weight for each time step: ， ， in represents the score function added to the attention model, and To learn the parameters, is a weight variable, T represents matrix transpose, represents the attention weight at time step t; Calculate the weighted output c: ； Step 3-2: The output c of step 3-1 is used as input into the physical information neural network PINN for optimization; Step 3-3: The output of step 3-2 is used as input into the knowledge enhancement network KAN for further optimization to obtain a preliminary prediction value; In step 3-4, the error sequence is obtained through the preliminary prediction value, and the error sequence is predicted using the bidirectional gated recurrent unit BiGRU. The predicted error sequence is arithmetic summed with the preliminary prediction result to obtain the final prediction result.

6. The method according to claim 5, characterized in that Step 3-2 includes: Step 3-2-1, build a neural network framework: build a neural network containing physical information ,in are the parameters of the network, is the predicted physical quantity, and They are spatial variables and time variables; neural network Able to receive the weighted output c from step 3-1 and incorporate geographical location factors and cloud cover, temperature and meteorological data; Step 3-2-2, integrating the laws of physics; Integrate the law of conservation of energy, the radiation transfer equation and the atmospheric radiation model and express it in the form of partial differential equations, where the law of conservation of energy is expressed as: ， in represents the rate of change of energy density, is the divergence of the energy flow; The radiation transfer equation is expressed as: ； in, represents the radiation intensity, represents the starting angle, is the radiation angle, s is the distance along the propagation direction, is the absorption coefficient, B is the blackbody radiation intensity, is the scattering phase function, which indicates the scattering angle The scattered light intensity in the direction, is the solid angle element; I is the irradiation; The atmospheric radiation model is expressed as: ， Among them, z represents the vertical height, represents the absorption coefficient, represents the scattering coefficient, represents the scattering phase function, represents the scattering angle, represents the scattering intensity, represents the atmospheric emission intensity at height z; Step 3-2-3, define the loss function; Assume that the partial differential equations for the radiation transfer equation and the atmospheric radiation model in step 3-2-2 are expressed as: ， Where L is an operator involving derivative operations; The residual of the partial differential equation in step 3-2-2 is used as part of the loss function, and the data loss, physical damage and regularization terms are defined separately; the loss function L is a weighted combination of the data loss and the residual loss of the partial differential equation; Assume the original GRU loss function is the mean square error , PDE residual loss Expressed as: ， ， in, is the weight coefficient; Step 3-2-5, training network: Use the loss function with physical information to train the neural network and optimize the data loss and physical loss to obtain the optimized prediction results. First, perform forward propagation and calculate the network output and PDE residuals ; Then, the loss function is obtained by weighted combination ; Finally, perform back propagation to calculate the network parameters The gradient of , and update the network parameters according to the gradient descent method , the expression is: ， in, represents the learning rate, Indicates network parameters The gradient, It means using the formula on the right to update the network parameters to obtain the preliminary predicted DNI sequence.

7. The method according to claim 6, characterized in that Step 3-3 includes: Step 3-3-1, knowledge extraction: extract relevant knowledge that affects the prediction of DNI time series data and encode the relevant knowledge into an operational mathematical form; Step 3-3-2, knowledge embedding: Let the original GRU loss function be the mean square error , the loss function after integrating knowledge is modified as follows: ， in, is the loss term related to knowledge, is the weight coefficient that balances data loss and knowledge loss, Defined as: ； in, is the output of the model being trained at the i-th sample point, is the output of the prior knowledge model in the GRU network at time i; Prediction results for: ， in, is the output of the GRU network, It is the weight to adjust the influence of knowledge.

8. An electronic device, characterized in that: The method comprises a processor and a memory, wherein the memory stores program codes, and when the program codes are executed by the processor, the processor is caused to perform the steps of the method according to any one of claims 1 to 7.

9. A storage medium, characterized in that: A computer program or instruction is stored, and when the computer program or instruction is run on a computer, the steps of the method according to any one of claims 1 to 7 are executed.

Images