A solar radiation and trough type photo-thermal system power generation prediction method for arid climate regions

Abstract

The present application relates to the technical field of solar power generation, and provides a solar radiation and trough photothermal system power generation prediction method for arid climate regions, first, a global horizontal total radiation sequence is decoupled in a frequency domain by using a discrete wavelet transform to extract a long-period component representing a physical trend; second, decoupled features are combined with multiple meteorological variables to be input into a time convolution network, time series deep fusion is carried out by using an expansion causal convolution to realize high-precision total radiation prediction of multiple steps; finally, the predicted total radiation is converted into direct normal radiation, which drives a quasi-steady physical model of a trough photothermal system to measure and calculate system net power generation. The present application can effectively eliminate random meteorological disturbances and reduce prediction errors, and establishes a closed loop from meteorological data prediction to photothermal physical system power generation estimation, which provides reliable technical support for grid connection scheduling and energy storage configuration of a trough photothermal power station in an arid climate region.

Description

Technical Field

[0001] This invention relates to the field of solar power generation technology, and in particular to a method for predicting the power generation of solar radiation and parabolic trough solar thermal systems in arid climate zones. Background Technology

[0002] Against the backdrop of "dual carbon" goals and energy transition, parabolic trough solar thermal power generation (PTC) systems occupy an important position in the renewable energy system due to their advantages of easy integration with energy storage and stable output. The concentrating solar power process of PTC is highly dependent on direct normal radiation (DNI). However, in actual meteorological monitoring systems, obtaining direct normal radiation is extremely difficult and costly. It is usually converted and measured using the more widely covered and easily accessible total horizontal radiation (GHI). The path of GHI to the Earth's surface is strongly interfered with by meteorological factors such as clouds and aerosols, exhibiting strong non-stationarity and multi-scale oscillations.

[0003] In recent years, deep learning techniques, represented by Temporal Convolutional Networks (TCNs) and Long Short-Term Memory Networks (LSTMs), have been widely applied to solar radiation prediction. However, most existing research remains at the level of meteorological data prediction, with very little exploration of the deep coupling effect between data-driven GHI predictions and physical models of solar thermal systems. The progressive impact of radiation prediction errors on actual PTC power generation assessment after the GHI to DNI conversion has not been systematically elucidated. Summary of the Invention

[0004] The purpose of this invention is to address the technical problems of insufficient prediction accuracy and ineffective integration of prediction results with solar thermal power generation physics systems in existing solar radiation forecasting methods. This invention provides a method for predicting solar radiation and power generation from parabolic trough solar thermal systems in arid climate zones. This method leverages the powerful inherent feature extraction capabilities of the TCN (Transient Catalytic Reduction) system to directly process multidimensional meteorological variables to predict total horizontal radiation. Furthermore, it transforms the high-precision predicted GHI (Gross Hierarchical Indicator) sequence into DNI (Digital Nitrogen Indicator) to drive the PTC (Physical Transformer) quasi-steady-state physics model, thus achieving a closed loop from radiation prediction to power generation assessment.

[0005] This invention provides the following technical solution:

[0006] A method for predicting solar radiation and power generation from parabolic trough solar thermal systems in arid climate zones includes the following steps:

[0007] Step 1: Obtain historical meteorological data and GHI sequences for the target area, preprocess the data, and construct time series features;

[0008] Step 2: Use discrete wavelet transform to decouple the GHI historical sequence in the frequency domain at multiple scales, and extract the low-frequency trend component that represents the daily periodic pattern and the high-frequency detail component that represents the meteorological random disturbance.

[0009] Step 3: Align the decoupled multi-scale components with other meteorological variables in the historical meteorological data in time series, and construct the input feature matrix using a sliding window;

[0010] Step 4: Input the input feature matrix into the TCN, and perform temporal deep fusion through dilated causal convolution and residual learning to predict the GHI for the next 1 to 4 hours;

[0011] Step 5: Convert the predicted GHI into DNI using an empirical model;

[0012] Step 6: Input the DNI into the quasi-steady-state physical model of PTC to calculate the net power generation of the system.

[0013] According to some possible implementation methods, the historical meteorological data in step one is used for feature screening through Pearson correlation coefficient, and features with an absolute value of correlation coefficient greater than 0.15 are retained as model input.

[0014] According to some possible implementations, the discrete wavelet transform in step two uses the Daubechies 4 wavelet as the basis wavelet to perform a two-level discrete decomposition on the GHI sequence. The decomposition expression is as follows:

[0015] GHI(t)=A2(t)+D2(t)+D1(t) (1)

[0016] Wherein, A2(t) is the low-frequency trend component characterizing the daily cycle pattern, and D2(t) and D1(t) are high-frequency detail components at different scales.

[0017] According to some possible implementation methods, the low-frequency trend component in step two is verified by the autocorrelation function at the 24-hour lag point. When the autocorrelation coefficient of the low-frequency trend component at the 24-hour lag point is higher than the autocorrelation coefficient of the original sequence, the decoupling is confirmed to be effective.

[0018] According to some possible implementations, the length of the sliding window in step three is set to 24 hours, with multidimensional time series data from the past 24 hours as model input and GHI from 1 to 4 hours after the end of the window as the prediction target.

[0019] According to some possible implementations, the formula for calculating the dilated causal convolution of the temporal convolutional network in step four is as follows:

[0020] (2)

[0021] Where d is the dilation factor, k is the kernel size, x is the input sequence, and f is the kernel.

[0022] According to some possible implementation methods, the radiation separation model in step five is the Erbs empirical model, which dynamically separates GHI into scattered horizontal radiation and DNI using the clear sky index. The calculation formula is as follows:

[0023] (3)

[0024] in, is the coefficient of the scattered radiation component.

[0025] According to some possible implementations, the quasi-steady-state physical model of the trough-type solar thermal system in step six includes an incident angle correction factor. This incident angle correction factor is determined after calculating the actual solar ray incident angle based on the solar declination angle and solar hour angle. It is used to correct optical loss caused by non-normal incident light, and the calculation formula is as follows:

[0026] (4)

[0027] (5)

[0028] (6)

[0029] (7)

[0030] in, This represents the instantaneous net power generation of the PT system; The total effective light-receiving area of ​​the solar collector field. Represents the dynamic optical efficiency of the system; The nominal maximum optical efficiency when sunlight is directly incident is taken as 0.75; and These are the system's pipe heat loss efficiency and the power island thermoelectric conversion efficiency, respectively. and These represent the solar declination angle and the solar hour angle, respectively. This is the actual angle of incidence of sunlight; This is the incident angle correction factor, used to correct for the additional optical loss caused by non-normal incident light.

[0031] According to some possible implementations, the calculation of net power generation in step six also involves the total effective light-receiving area of ​​the collector field, the dynamic optical efficiency of the system, the heat loss efficiency of the pipeline, and the thermoelectric conversion efficiency parameters of the power island.

[0032] Compared with the prior art, the present invention has the following beneficial effects:

[0033] This invention introduces discrete wavelet transform to decouple the non-stationary horizontal total radiation sequence in the multi-scale frequency domain, effectively separating the stable low-frequency component reflecting the astronomical cycle from the high-frequency component reflecting meteorological random disturbances. Autocorrelation function analysis confirms that the decoupled low-frequency trend term (A2) completely preserves the 24-hour diurnal physical cycle regularity of the original sequence. By constructing a quasi-steady-state physical model of the parabolic trough solar thermal system, it breaks through the limitation of existing research that only focuses on meteorological data prediction. By dynamically converting high-precision GHI prediction values ​​into DNI through the Erbs empirical model, and then driving the quasi-steady-state physical model of parabolic trough solar thermal power generation including the incident angle correction factor, it realizes the direct calculation of the system's net power generation. This method provides a reliable engineering foundation for the grid-connected integration and real-time scheduling of parabolic trough solar thermal power plants. Attached Figure Description

[0034] Figure 1 This is a schematic diagram of the overall process of the solar radiation and power generation prediction method for trough solar thermal systems in arid climate zones provided by an embodiment of the present invention.

[0035] Figure 2 This is a correlation analysis diagram between meteorological characteristics and GHI provided in an embodiment of the present invention.

[0036] Figure 3 This is a flowchart of the conversion estimation of solar radiation to net PTC power generation provided in an embodiment of the present invention.

[0037] Figure 4 This is an ACF analysis diagram of the original solar radiation sequence and its decoupled components (A2, D2) provided in the embodiments of the present invention.

[0038] Figure 5 This is a comparison chart of prediction performance under different prediction step sizes provided in the embodiments of the present invention.

[0039] Figure 6 This is a comparison chart of the prediction performance of the TCN and W-TCN models provided in the embodiments of the present invention.

[0040] Figure 7 This is a performance comparison chart of different models provided in the embodiments of the present invention in different seasons.

[0041] Figure 8 This is a comparison chart of power generation of a parabolic trough solar thermal system based on real and predicted solar radiation, provided by an embodiment of the present invention.

[0042] Figure 9 Table 1 is provided in the embodiments of the present invention. Table 1 is a table of monthly average error results.

[0043] Figure 10 Table 2 is provided in the embodiments of the present invention. Table 2 is a table of hyperparameter setting ranges for each model. Detailed Implementation

[0044] The present invention will now be described in detail with reference to embodiments and accompanying drawings. However, it should be understood that the embodiments and drawings are for illustrative purposes only and do not constitute any limitation on the scope of protection of the present invention. All reasonable modifications and combinations included within the inventive spirit of the present invention fall within the scope of protection of the present invention.

[0045] Example 1

[0046] This embodiment provides the implementation process of a method for predicting solar radiation and power generation of parabolic trough solar thermal systems in arid climate zones, such as... Figure 1 As shown, it includes the following steps:

[0047] Step 1: Obtain historical meteorological data and GHI sequences for the target area, preprocess the data, and construct time series features.

[0048] The data source is the ERA5 reanalysis dataset released by the European Centre for Medium-Range Weather Forecasts (ECMWF). Multiple features were extracted from the original dataset, including month, day, hour, temperature, sea level pressure, humidity, wind speed, saturated vapor pressure, daytime / nighttime status, radiation duration, GHI, dew point temperature, rainfall, atmospheric pressure, wind direction, and cloud cover. Feature selection was performed to improve model efficiency and suppress noise. (See attached image.) Figure 2 As shown, the Pearson correlation coefficient was used to quantitatively assess the linear relationship between each meteorological variable and the GHI.

[0049] Step 2: Use discrete wavelet transform to decouple the GHI historical sequence in the frequency domain at multiple scales, and extract the low-frequency trend component that represents the daily periodic pattern and the high-frequency detail component that represents the meteorological random disturbance.

[0050] Discrete wavelet transform is employed to decouple the GHI historical sequence across multiple scales in the frequency domain. Multi-resolution analysis of the radiative sequence is achieved through scaling and shifting wavelet basis functions. This invention uses Daubechies 4 (db4), which exhibits good compact support and approximate symmetry, as the basis wavelet, and employs the Mallat pyramid algorithm to perform a two-level discrete decomposition of the original radiative sequence. Figure 4 As shown, the ACF of the original radiation sequence exhibits significant peaks at lag points such as 24 and 48 hours, demonstrating a stable diurnal cycle. After DWT decomposition, the peak value of the ACF curve of the low-frequency component A2 at the 24-hour lag point is significantly enhanced, and it maintains a more complete sinusoidal fluctuation pattern throughout the entire lag interval, with a clearer periodic structure than the original sequence.

[0051] Step 3: Align the decoupled multi-scale components with other meteorological variables in the historical meteorological data in time series, and construct the input feature matrix using a sliding window.

[0052] The sliding window method is used to construct time series features. The window length is set to 24 hours. That is, the multidimensional data of the past 24 hours (including the original meteorological features and the decoupled A2, D2, D1 components) are used as the model input, and the GHI of the following 1, 2, 3, and 4 hours after the end of the window is used as the prediction target.

[0053] Step 4: Input the input feature matrix into a temporal convolutional network, and perform temporal deep fusion through dilated causal convolution and residual learning to predict GHI for the next 1 to 4 hours.

[0054] The input feature matrix is ​​fed into a temporal convolutional network. The core of the W-TCN model constructed in this invention lies in the high degree of fit between the decoupled multi-scale features and the TCN network topology. Causal convolution is used to ensure no leakage of future information, combined with dilated convolution to expand the receptive field. For example... Figure 5 As shown, as the prediction step size increases from 1 hour to 4 hours, the prediction performance of all models degrades to some extent. However, W-TCN maintains a low prediction error across all step sizes, and the error increases with the smallest increment. Figure 6 and Figure 7 As shown, the predicted values ​​of W-TCN on the test set are closely distributed near the diagonal of the true values, and the density curves of the marginal histograms highly overlap. Figure 9 As shown in Table 1, the monthly average MAE error results of W-TCN under various months and prediction step sizes verify the stable prediction performance of the model under different seasonal conditions. In terms of seasonality, the prediction uncertainty of all models increases in summer, but W-TCN maintains the best performance in all seasons, especially in summer when radiation fluctuations are strong, and its ability to capture peaks is significantly better than other models.

[0055] Step 5: Convert the predicted GHI into DNI using an empirical model.

[0056] Using the classic Erbs empirical model, and through the clear sky index K... t The scattered radiation component coefficients are dynamically calculated. The predicted GHI is then separated to obtain the DNI.

[0057] Step 6: Input the DNI into the quasi-steady-state physical model of PTC to calculate the net power generation of the system.

[0058] like Figure 3 As shown, the calculated DNI is input into the quasi-steady-state physical model of the PTC. This model fully considers the optical losses when light is not incident normally, includes an incident angle correction factor, and yields the power generation. Figure 8As shown, this invention converts the GHI predicted based on W-TCN into DNI, and then drives the PTC quasi-steady-state physical model to obtain a net power generation curve that maintains a high degree of consistency with the overall trend of the actual net power generation curve.

[0059] Example 2

[0060] This embodiment serves as a comparative example, providing a comparative experiment without introducing discrete wavelet transform for frequency domain decoupling to highlight the advantages of the present invention. To obtain the optimal configuration for each model, the Optuna framework is used for automatic hyperparameter optimization, with minimizing the prediction error on the validation set as the optimization objective. The hyperparameter search space for each model is as follows: Figure 10 As shown in Table 2.

[0061] The implementation process of this comparative example is basically the same as that of Example 1. The difference is that the discrete wavelet transform operation is omitted in step two. The original GHI sequence is directly time-aligned with the historical meteorological features selected in step one and then input into the TCN model for prediction. The remaining steps (including Erbs model radiation separation and PTC quasi-steady-state physical model driving) remain unchanged.

[0062] Experimental results show that, without introducing DWT frequency domain decoupling, the MAE of the TCN model at prediction step sizes of 1 hour, 2 hours, 3 hours, and 4 hours in Lanzhou is 17.26 W / m², 25.80 W / m², 31.03 W / m², and 35.82 W / m², respectively. In contrast, the MAE of the W-TCN model (Example 1) after introducing DWT decoupling decreases to 6.24 W / m², 12.51 W / m², 17.08 W / m², and 25.08 W / m² at each step size, representing a significant improvement. Furthermore, without introducing DWT, the model's prediction deviation for peak radiation increases significantly during summer months with drastic cloud cover fluctuations (such as June and August), indicating that high-frequency random disturbances severely interfere with the model's prediction accuracy. This comparative example fully verifies the necessity and effectiveness of the discrete wavelet transform frequency domain decoupling operation in this invention.

[0063] The above embodiments are merely preferred embodiments of the present invention, and the scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A method for predicting solar radiation and power generation of parabolic trough solar thermal systems in arid climate zones, characterized in that, Includes the following steps: Step 1: Obtain historical meteorological data and GHI sequences for the target area, preprocess the data, and construct time series features; Step 2: Use discrete wavelet transform to decouple the GHI historical sequence in the frequency domain at multiple scales, and extract the low-frequency trend component that represents the daily periodic pattern and the high-frequency detail component that represents the meteorological random disturbance. Step 3: Align the decoupled multi-scale components with other meteorological variables in the historical meteorological data in time series, and construct the input feature matrix using a sliding window; Step 4: Input the input feature matrix into the TCN, and perform temporal deep fusion through dilated causal convolution and residual learning to predict the GHI for the next 1 to 4 hours; Step 5: Convert the predicted GHI into DNI using an empirical model; Step 6: Input the DNI into the quasi-steady-state physical model of PTC to calculate the net power generation of the system.

2. The prediction method according to claim 1, characterized in that, The historical meteorological data mentioned in step one is used for feature filtering through Pearson correlation coefficient, and features with an absolute value of correlation coefficient greater than 0.15 are retained as model input.

3. The prediction method according to claim 2, characterized in that, The discrete wavelet transform described in step two uses the Daubechies 4 wavelet as the basis wavelet to perform a two-level discrete decomposition on the GHI sequence. The decomposition expression is as follows: GHI(t)=A2(t)+D2(t)+D1(t) (1) Wherein, A2(t) is the low-frequency trend component characterizing the daily cycle pattern, and D2(t) and D1(t) are high-frequency detail components at different scales.

4. The prediction method according to claim 3, characterized in that, In step two, the low-frequency trend component is verified at the 24-hour lag point using the autocorrelation function. When the autocorrelation coefficient of the low-frequency trend component at the 24-hour lag point is higher than that of the original sequence, the decoupling is confirmed to be effective.

5. The prediction method according to claim 4, characterized in that, In step three, the length of the sliding window is set to 24 hours, using multidimensional time-series data from the past 24 hours as model input, and GHI from 1 to 4 hours after the end of the window as the prediction target.

6. The prediction method according to claim 5, characterized in that, The formula for calculating the dilated causal convolution of the temporal convolutional network described in step four is as follows: （2） Where d is the dilation factor, k is the kernel size, x is the input sequence, and f is the kernel.

7. The prediction method according to claim 6, characterized in that, The radiation separation model described in step five is the Erbs empirical model, which dynamically separates the GHI into scattered horizontal radiation and DNI using the clear sky index. The calculation formula is as follows: （3） in, is the coefficient of the scattered radiation component.

8. The prediction method according to claim 7, characterized in that, The quasi-steady-state physical model of the trough-type solar thermal system described in step six includes an incident angle correction factor. This factor is determined by calculating the actual incident angle of sunlight based on the solar declination angle and solar hour angle. It is used to correct for optical loss caused by non-normal incident light. The calculation formula is as follows: （4） （5） （6） （7） in, This represents the instantaneous net power generation of the PT system; The total effective light-receiving area of ​​the solar collector field. Represents the dynamic optical efficiency of the system; The nominal maximum optical efficiency when sunlight is directly incident is taken as 0.75; and These are the system's pipe heat loss efficiency and the power island thermoelectric conversion efficiency, respectively. and These represent the solar declination angle and the solar hour angle, respectively. This is the actual angle of incidence of sunlight; This is the incident angle correction factor, used to correct for the additional optical loss caused by non-normal incident light.

9. The prediction method according to claim 2, characterized in that, The calculation of net power generation in step six also involves the total effective light-receiving area of ​​the solar collector field, the dynamic optical efficiency of the system, the heat loss efficiency of the pipeline, and the thermoelectric conversion efficiency parameters of the power island.

Images