Ultra-short-term irradiation prediction method for photovoltaic power stations based on single-photon lidar

Abstract

A method for ultra-short-term irradiation prediction for photovoltaic power plants based on single-photon lidar belongs to the field of lidar signal processing technology. It solves the problem of how to use lidar to obtain cloud information and realize ultra-short-term irradiation prediction for photovoltaic power plants under different weather conditions. The lidar echo signal profile is inverted by an improved differential zero-crossing algorithm to obtain a cloud parameter dataset. Based on the volume scanning mode of the radar, the cloud situation in the sky can be inverted to obtain more characteristic cloud parameters, ensuring the integrity and accuracy of the data. The sun path is further calculated, the effective scanning area is segmented to extract cloud features and reduce the feature dimension. Finally, a feature matrix is ​​constructed by combining meteorological characteristics and historical irradiation data. The irradiation is predicted using a CNN-LSTM hybrid model. The predictions show high accuracy and stability in different weather types.

Description

Technical Field

[0001] The present invention belongs to the technical field of laser radar signal processing and relates to an ultra-short-term irradiation prediction method for a photovoltaic power station based on a single-photon laser radar. Background Art

[0002] Amidst increasing energy shortages and environmental pressures, new energy sources are playing an increasingly important role in the global energy mix, with photovoltaic power generation becoming a core component of renewable energy. As the primary driver of photovoltaic power generation, accurate prediction of solar irradiance is crucial for optimizing grid scheduling, improving the operational efficiency of photovoltaic power plants, and achieving sustainable energy development. However, solar irradiance is affected by weather conditions, particularly cloud movement, and exhibits frequent and dramatic fluctuations, resulting in significant volatility in photovoltaic power generation. Accurately predicting irradiance for photovoltaic power plants is crucial to effectively address this issue.

[0003] Existing irradiation prediction methods can be divided into ultra-short-term, short-term, and medium- to long-term predictions according to the different prediction time periods. Ultra-short-term predictions focus on irradiation fluctuations within a few hours to a day, and have high timeliness and accuracy requirements. At present, ultra-short-term irradiation prediction methods can be mainly divided into the following three categories: (1) clear sky models derived from empirical formulas of surface solar irradiance; derivation models based on physical models and empirical formulas have simple inputs, such as the ASHRAE model, Hottel model, REST2 model, and Iqbal-C model. However, traditional prediction methods based on physical models have low accuracy under complex weather conditions and are difficult to handle nonlinear and multi-dimensional meteorological influences. (2) Time series prediction methods based on historical data and numerical weather forecasts (NWP); There are two research trends in the methods of machine learning or deep learning prediction models based on historical data and NWP. One is the development of regression models, such as the autoregressive integrated moving average model (ARIMA), which is a classic time series method, but it relies on NWP information. In order to achieve higher prediction accuracy, more and more detailed meteorological information needs to be input; the other is the development of time series prediction models that rely on historical data and learn trend changes from more historical data, such as long short-term memory networks (LSTM), convolutional neural networks (CNN), Transformer, Informer and other models. Although existing technologies have achieved certain results in irradiation prediction, there are still some shortcomings. Although machine learning models can handle nonlinear problems well, their training process is complex and relies heavily on data quality. If the data is insufficient or contains noise, it will seriously affect the prediction effect of the model. (3) Prediction methods combined with cloud maps; Since cloud layer is the main factor affecting irradiance, irradiance prediction combined with cloud changes is still a research hotspot.

[0004] Currently, cloud characteristics are primarily acquired from satellite and ground-based cloud images. Ground-based cloud images are images of clouds above the station being measured, collected in real time by ground-based all-sky imagers. Satellite cloud images are obtained by meteorological satellites observing the Earth as a whole. In addition to satellites and optical cameras, cloud observation instruments also include lidar (LiDAR). As an advanced meteorological observation tool, LiDAR can acquire atmospheric physics information by emitting laser pulses and receiving the scattered echo signals in the atmosphere. It has high spatial and temporal resolution, enabling precise detection of cloud distribution and changes, providing valuable data support for solar irradiance prediction. Prior art, such as the invention patent application with publication number CN114296050A, discloses a method for predicting the short-term power generation of photovoltaic power stations based on LiDAR cloud imagery. This method divides the effective area of ​​the photovoltaic power station's solar panels into grids. Based on the cloud boundary range and movement speed, the predicted values ​​of the cloud shadow's residence time in each grid and the ratio of the cloud shadow's area to each grid are calculated, thereby deriving the predicted photovoltaic power generation value. However, this method only considers the position and motion prediction of clouds and cannot accurately evaluate the impact of different types of clouds on irradiance, and thus on power.

[0005] Existing ultra-short-term irradiance prediction methods primarily include those based on satellite cloud images and those based on ground-based cloud images. Satellite cloud images typically have low resolution, and may not provide sufficient detail, particularly in areas with thin or light cloud cover. Satellite cloud images are acquired infrequently, typically only once every hour or several hours. This results in an inability to provide real-time information on ultra-short-term irradiance changes. In dense cloud conditions, satellites may not be able to clearly observe the ground, compromising the integrity and accuracy of the data. Ground-based cloud images primarily come from all-sky imagers. While they have a wide viewing angle, they still cannot cover all areas on the ground. They may have blind spots, especially in areas with tall buildings or complex terrain. Image quality is also affected by environmental factors such as atmospheric pollution, varying lighting conditions, and haze, which can cause image blur or interference, impacting cloud analysis and prediction accuracy. Furthermore, processing all-sky images requires extensive computation, including image preprocessing, cloud recognition, and feature extraction. This results in high computational complexity and demands for high computing resources. Summary of the Invention

[0006] The technical solution of the present invention is used to solve the problem of how to use lidar to obtain cloud information and realize ultra-short-term irradiation prediction of photovoltaic power stations under different weather conditions.

[0007] The present invention solves the above technical problems through the following technical solutions:

[0008] The ultra-short-term irradiation prediction method for photovoltaic power plants based on single-photon lidar includes the following steps:

[0009] Step S1: Setting the scanning mode of the single-photon laser radar to a volume scanning mode, scanning the sky near the photovoltaic power station at a fixed elevation angle to obtain echo signals in different directions;

[0010] Step S2: For each echo signal acquired by the laser radar scan, select the echo signal within 50-100 meters of the highest detection target distance, use its mean as the background noise, and subtract the background noise from the original echo signal to obtain the corrected echo signal;

[0011] Step S3, using the differential zero-crossing method to identify cloud information in each echo signal;

[0012] Step S4, obtaining scanning cycle data obtained by rotating the laser radar within the azimuth range, calculating the variation range of the solar azimuth angle of the laser radar within one scanning cycle, and intercepting the key area data according to the time point closest to the solar azimuth angle in the scanning cycle to obtain an intercepted data set;

[0013] Step S5: reducing the data dimension and extracting more accurate cloud features;

[0014] Step S6: Select meteorological data and historical irradiation data recorded by the meteorological monitoring platform of the photovoltaic power station, delete missing data and erroneous data, obtain historical time-series irradiation and time-series meteorological characteristics, merge the historical time-series irradiation, time-series meteorological characteristics, and the cloud layer characteristics extracted in step S5 into all feature vectors at that moment, and intercept them using a fixed time interval sliding window method to obtain an input feature matrix;

[0015] Step S7: construct a CNN-LSTM model, input the feature matrix into the convolutional neural network for feature extraction, input the extracted time series features into the long short-term memory layer for irradiance prediction, and obtain the irradiance value for the time period to be predicted.

[0016] Furthermore, the step S1 includes the following steps:

[0017] Step S11, setting the elevation angle of the laser radar to a fixed value. At the fixed elevation angle, the laser radar rotates around the current azimuth angle α at a constant rate;

[0018] Step S12: After the laser radar emits a single-photon laser pulse at the current azimuth angle α, the laser radar receives an echo signal P from the direction. α (r i ), the signal strength received by the single-scan laser radar is Among them, r1, r2, ..., r NIndicates the distance of the detected target, N indicates the total number of echo signal sampling points, and i indicates the i-th echo signal sampling point;

[0019] In step S13, after completing the signal acquisition at the current azimuth angle α, the laser radar continues to change the azimuth angle, emitting single-photon laser pulses and receiving echo signals to the next azimuth angle α+Δα until the predetermined azimuth angle range is covered.

[0020] Furthermore, step S3 includes the following steps:

[0021] Step S31: distance correction is performed on the original echo signal to eliminate signal attenuation due to distance attenuation. The following formula is used to perform distance correction on the original echo signal:

[0022] X α (r) = P′ α (r)r 2

[0023] Where P′ α (r) represents the backscattered signal of the lidar at distance r, X α (r) represents the value of P′ α (r) The signal after distance square correction, r represents the distance of the detected target;

[0024] Step S32, differentiating the pre-processed echo signal and fitting the differential signal using a sliding window;

[0025] Step S33: Determine the cloud base and cloud peak positions in the echo signal profile using the differential zero-crossing method to determine the cloud base signal r b 、Yunfeng Signal m 、Genting Signal t , preliminarily screen out candidate signal areas containing cloud signals;

[0026] Step S34 , screening cloud areas using a threshold value, eliminating misjudgments caused by thick aerosols or background light based on the signal intensity ratio, and determining the number of clouds.

[0027] Furthermore, the step S34 uses the following formula to filter the cloud area:

[0028]

[0029] Where thr represents the threshold, X α (r m ) represents the cloud peak height, X α (r b ) represents the cloud base height, P′(r m ) represents the distance from the cloud peak signal r mThe backscattered signal of the laser radar at b ) represents the distance from the cloud base signal r b The lidar backscatter signal at is judged to contain cloud signals when the threshold is greater than 4, and is judged to contain aerosol signals when the threshold is less than or equal to 4.

[0030] Furthermore, step S4 includes the following steps:

[0031] Step S41, calculating the variation range of the solar azimuth angle of the laser radar within one scanning cycle;

[0032] Step S42, within a scanning cycle of the laser radar, based on the time point corresponding to the closest solar azimuth angle in the scanning cycle, intercept the scanning data of the adjacent 30 seconds, a total of 1 minute of data.

[0033] Furthermore, step S41 calculates the variation range of the solar azimuth angle of the laser radar within one scanning cycle using the following formula:

[0034]

[0035] H=15°×(T-12)

[0036]

[0037] Where δ represents the solar declination angle, n represents the nth day of the year, H represents the solar hour angle, T represents the solar time, φ represents the geographical latitude, and α represents the solar hour angle. sun Indicates the solar azimuth.

[0038] Furthermore, the cloud characteristics in step S5 include solar altitude angle, solar azimuth angle, periodic cloud fraction, cloud base height, cloud thickness, number of cloud layers, and regional cloud fraction.

[0039] Furthermore, step S5 includes the following steps:

[0040] Step S51: analyzing the scan cycle data obtained by rotating the lidar within the azimuth range according to step S3, inverting each lidar echo signal to determine whether there is a cloud signal. If there is a cloud signal, determining the cloud base height, cloud thickness, and cloud number;

[0041] Step S52: Count the number of echo signals containing cloud signals in the entire scanning cycle data, and compare it with the number of echo signals in the entire scanning cycle to obtain the cloud score in the entire scanning cycle, which is used to represent the cloud coverage ratio in the sky during the entire scanning cycle;

[0042] Step S53: extracting the cloud base height corresponding to the echo signal in the intercepted data set and calculating the mean of all cloud base heights in the key area; wherein, if there are multiple layers of clouds in the key area, the cloud base height of the lowest layer is taken as the representative value of the cloud base height in the area;

[0043] Step S54: extracting the mean cloud thickness in the key area; if there are multiple layers of cloud in the key area, the thickness of each layer of cloud is accumulated to obtain the total cloud thickness;

[0044] Step S55, counting the number of cloud layers obtained by inverting the echo signals in the key area, and analyzing the mode of the number of cloud layers in the area to characterize the number of cloud layers in the area;

[0045] Step S56 , counting the number of echo signals containing cloud signals in the key area, and comparing it with the number of echo signals in the entire key area, to obtain a regional cloud score in the key area.

[0046] Furthermore, the meteorological information in step S6 includes temperature, relative humidity, air pressure, and wind speed.

[0047] Furthermore, the CNN-LSTM model in step S7 includes an input layer, a convolutional layer, a pooling layer, an LSTM layer, a fully connected layer and an output layer.

[0048] The advantages of the present invention are:

[0049] (1) The present invention uses a single-photon lidar to scan the cloud information around the photovoltaic power station, detect key cloud parameters such as cloud coverage, cloud thickness, and cloud height in the area, extract cloud characteristics, combine historical irradiance and meteorological data, and use the CNN-LSTM time series prediction algorithm to achieve high-precision irradiance prediction.

[0050] (2) The present invention inverts the echo signal profile of the lidar through an improved differential zero-crossing algorithm to obtain a cloud parameter data set. Based on the volume scanning mode of the radar, the situation of clouds in the sky can be inverted. The radar scanning range is larger than that of the ground-based cloud map taken by the fisheye lens, and the data resolution is higher than that of the satellite cloud map. It can also obtain more characteristic parameters of the cloud, ensuring the integrity and accuracy of the data.

[0051] (3) The present invention further calculates the sun path through the sun position model, divides the effective scanning area to extract cloud features, reduces the feature dimension, and finally constructs a feature matrix by combining meteorological characteristics and historical irradiation data. The CNN-LSTM hybrid model is used to predict irradiation, showing high accuracy and stability in the prediction of different weather types. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1Schematic diagram of a method for predicting ultra-short-term irradiation of a photovoltaic power station based on a single-photon lidar according to a first embodiment of the present invention;

[0053] Figure 2 Schematic diagram of a single-photon laser radar in a volume scanning mode according to a first embodiment of the present invention;

[0054] Figure 3 1 is a schematic diagram of a differential zero-crossing algorithm according to a first embodiment of the present invention;

[0055] Figure 4 Schematic diagram of data heat and area capture of the single-photon laser radar in one scanning cycle according to the first embodiment of the present invention;

[0056] Figure 5 This is a schematic diagram of the CNN-LSTM model of the first embodiment of the present invention. DETAILED DESCRIPTION

[0057] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0058] The technical solution of the present invention is further described below with reference to the accompanying drawings and specific embodiments:

[0059] Example 1

[0060] like Figure 1 Specifically, a method for predicting ultra-short-term irradiation of a photovoltaic power station based on a single-photon laser radar is disclosed, comprising the following steps:

[0061] Step S1: Set the scanning mode of the single-photon laser radar to volume scanning mode, scan the sky near the photovoltaic power station at a fixed elevation angle, and obtain echo signals in different directions. Specifically, step S1 includes the following steps:

[0062] Step S11, setting the elevation angle of the laser radar to a fixed value. At the fixed elevation angle, the laser radar rotates around the current azimuth angle α at a constant rate;

[0063] Step S12: After the laser radar emits a single-photon laser pulse at each azimuth angle, the laser radar receives an echo signal from that direction;

[0064] In this embodiment, after the laser radar emits a single-photon laser pulse at the current azimuth angle α, the laser radar receives an echo signal P from this direction.α (r i ), the signal strength received by the single-scan laser radar is The echo signal is a sequence set that decays with distance at different azimuth angles in space, where r1, r2, ..., r N Indicates the distance to the detected target, N indicates the total number of echo signal sampling points, and i indicates the i-th echo signal sampling point.

[0065] In step S13, after completing the signal acquisition at the current azimuth angle α, the laser radar continues to change the azimuth angle, emitting single-photon laser pulses and receiving echo signals to the next azimuth angle α+Δα until the predetermined azimuth angle range is covered, which is specifically 0-360°.

[0066] In existing technologies, lidar typically uses two modes: horizontal scanning (PPI) and vertical elevation scanning (RHI). The vertical elevation scanning (RHI) is primarily used to detect the vertical structure of the atmosphere, such as cloud height, aerosol layer thickness, and vertical profiles of temperature and humidity. The horizontal scanning (PPI) is suitable for detecting the horizontal distribution of aerosols and clouds over a large area at a fixed altitude, or for detecting local atmospheric phenomena within a smaller azimuth range.

[0067] However, because the azimuth and elevation of the sun change continuously throughout the day, the impact of clouds on irradiance must take into account the dynamics of the clouds around the sun. Traditional scanning modes usually scan at a fixed azimuth or set the elevation angle to 0°. This method is difficult to meet the cloud information detection required for irradiance prediction. The volume scanning mode (VCP) has significant advantages over traditional scanning methods. It can not only detect cloud distribution at different vertical heights, but also obtain cloud information within different azimuth ranges. Therefore, volume scanning can provide cloud data over a larger spatial range, greatly improving the accuracy of irradiance prediction.

[0068] like Figure 2 As shown, this embodiment uses a volumetric scanning mode (VCP) specifically designed to track cloud dynamics along the sun's path, rather than passively detecting clouds at fixed angles. Compared to existing all-sky imagers, lidar captures richer and more detailed cloud information through echo signals, providing stronger support for irradiance prediction.

[0069] Step S2: for each echo signal P obtained by the laser radar scanning α (r i ), select the echo signal within 50-100 meters of the highest detection target distance, take its mean as the background noise, and subtract the background noise from the original echo signal as the corrected echo signal. The corrected echo signal P′ is calculated using the following formula:α (r i ):

[0070]

[0071] Where P′ α (r i ) represents the corrected echo signal, and m represents the number of sampling points within 50-100 meters.

[0072] Step S3, using the differential zero-crossing method to identify the cloud information in each echo signal. Specifically, step S3 includes the following steps:

[0073] Step S31: pre-process the echo signal; specifically, perform distance correction on the original echo signal to eliminate signal attenuation due to distance attenuation, wherein the distance correction is performed on the original echo signal using the following formula:

[0074] X α (r) = P′ α (r)r 2

[0075] Where P′ α (r) represents the backscattered signal of the lidar at distance r, X α (r) represents the value of P′ α (r) The signal after distance square correction, r represents the distance of the detected target.

[0076] Step S32: Differentiate the preprocessed echo signal and fit the differential signal using a sliding window. Specifically, fitting the differential signal using a sliding window is performing 7-point sliding window fitting on the differential signal.

[0077] When using the differential zero-crossing method, the calculated differential signal (i.e., the rate of change of the echo signal) can have excessive zero crossings due to noise or data instability. These extra zero crossings, or zero crossings, do not represent actual physical phenomena but are caused by factors such as noise in the echo signal or sampling errors.

[0078] In this embodiment, in order to further reduce the zero-crossing points in the echo signal change rate (i.e., the differential signal) in the differential calculation, a 7-point sliding window is used to fit the differential signal. The signal is locally smoothed by using a fixed-size window. The window slides on the differential signal, and the data within the window is fitted each time it slides, thereby smoothing the signal fluctuations and reducing the impact of high-frequency noise on the signal.

[0079] Step S33, invert the cloud parameters; specifically, use the differential zero crossing method to determine the positions of the cloud base and cloud peak in the echo signal profile, and determine the cloud base signal rb 、Yunfeng Signal m 、Genting Signal t , preliminarily screen out candidate signal areas containing cloud signals;

[0080] Step S34, screening the clouds; specifically, screening the cloud area by using a threshold, eliminating misjudgments caused by thick aerosols or background light based on the signal intensity ratio, and determining the number of clouds;

[0081] The cloud area is screened using the following formula:

[0082]

[0083] Where thr represents the threshold, X α (r m ) represents the cloud peak height, X α (r b ) represents the cloud base height, P′(r m ) represents the distance from the cloud peak signal r m The backscattered signal of the laser radar at b ) represents the distance from the cloud base signal r b The laser radar backscatter signal at . This embodiment uses a threshold value thr = 4 to determine the cloud area. When the threshold value is greater than 4, it is determined to contain cloud signals. When the threshold value is less than or equal to 4, it is determined to be aerosol.

[0084] In this embodiment, due to differences in cloud droplet concentration, water content, and phase, the backscattered echo intensity also varies. This embodiment uses the differential zero-crossing method based on the important observation that aerosols are typically less dense than clouds, and the echo signal is attenuated before clouds are detected.

[0085] like Figure 3 As shown, when the laser hits the cloud boundary, its echo signal intensity P α (r i ) will surge. According to the characteristics of the laser radar echo signal, the position where its slope changes from negative to positive is the cloud bottom signal r b , represents the cloud base position of the cloud layer; when the signal strength reaches the maximum value and begins to decay, the position at this time is the cloud peak signal r of this cloud layer m , that is, the position where the slope of the lidar profile changes from positive to negative, indicating the position of the cloud peak.

[0086] When the laser passes through the cloud, P α (r i ) decreases rapidly, generating a cloud echo signal, r tIndicates the cloud top position. A cloud lidar echo signal can be expressed as a set of adjacent peaks and troughs on its corresponding first-order differential signal, while the cloud bottom signal r b 、Yunfeng Signal m 、Genting Signal t That is the zero point where its differential signal intersects the height coordinate axis.

[0087] Step S4, cropping the scanning area taking into account the path of the incident light; obtaining the scanning cycle data obtained by rotating the lidar within the azimuth range according to step S1, calculating the range of change of the solar azimuth angle of the lidar within one scanning cycle, and intercepting the key area data according to the time point closest to the solar azimuth angle in the scanning cycle to obtain a intercepted data set.

[0088] Specifically, the following steps are included:

[0089] like Figure 4 As shown, based on the volume scanning mode set by the single-photon laser radar in step S1, the scanning cycle data obtained by rotating the laser radar in the azimuth angle range of 0-360° can be obtained.

[0090] Step S41, calculate the sun path; specifically, use the following formula to calculate the range of change of the solar azimuth angle (i.e., the azimuth angle of incident sunlight) of the laser radar within one scanning cycle:

[0091]

[0092] H=15°×(T-12)

[0093]

[0094] Where δ represents the solar declination angle, n represents the nth day of the year, H represents the solar hour angle, T represents the solar time, φ represents the geographical latitude, and α represents the solar hour angle. sun Indicates the solar azimuth.

[0095] like Figure 4 As shown, in this embodiment, the scanning period is 11:58-13:07. Combined with the geographical location, date and time range of the laser radar installation location, the above formula is used to calculate the variation range of the solar azimuth angle during this time period:

[0096] Step S42, determine the key azimuth angle and intercept the data; within a scanning cycle of the lidar, the scanning azimuth angle of the lidar will almost overlap with the solar azimuth angle at a certain moment. The azimuth angle at this time is the key to extracting cloud features. According to the time point corresponding to the closest solar azimuth angle in the scanning cycle, intercept the scanning data of the adjacent 30 seconds, a total of 1 minute of data.

[0097] like Figure 4As shown in , during this scanning cycle, the closest angle between the solar azimuth and the lidar azimuth is 199.42°. The scanning data of the time point corresponding to this azimuth angle for 30 seconds is intercepted, a total of 1 minute of data, as shown in Figure 4 The range selected by the rectangular frame is shown in the figure.

[0098] Step S5: Reduce the data dimension and extract more accurate cloud features; the cloud features include solar altitude angle, solar azimuth angle, periodic cloud fraction, cloud base height, cloud thickness, number of cloud layers, and regional cloud fraction; specifically, step S5 includes the following steps:

[0099] Step S51: Analyze the scanning cycle data obtained by rotating the laser radar within the azimuth range according to step S3, invert each laser radar echo signal, and determine whether there is a cloud signal. If there is a cloud signal, determine the cloud base height, cloud thickness, and number of clouds.

[0100] Step S52: Count the number of echo signals containing cloud signals in the entire scanning cycle data, and compare it with the number of echo signals in the entire scanning cycle to obtain the cloud score in the entire scanning cycle, which is used to represent the cloud coverage ratio in the sky during the entire scanning cycle.

[0101] Step S53: extracting the cloud base height corresponding to the echo signal in the intercepted data set and calculating the mean of all cloud base heights in the key area;

[0102] Furthermore, if there are multiple layers of clouds in the key area, the cloud base height of the lowest layer is taken as the representative value of the cloud base height in the area.

[0103] Step S54, extracting the mean value of cloud thickness in the key area;

[0104] Furthermore, if there are multiple layers of clouds in the key area, the thickness of each cloud layer is accumulated to obtain the total cloud thickness.

[0105] Step S55 , counting the number of cloud layers obtained by inverting the echo signals in the key area, analyzing the mode of the number of cloud layers in the area, and using it to characterize the number of cloud layers in the area.

[0106] Step S56 , counting the number of echo signals containing cloud signals in the key area, and comparing it with the number of echo signals in the entire key area, to obtain a regional cloud score in the key area.

[0107] Step S6: Construct a feature matrix. Select meteorological data and historical irradiance data recorded by the PV power station's meteorological monitoring platform, delete any missing or erroneous data, and obtain historical time-series irradiance and time-series meteorological features. Combine the historical time-series irradiance, time-series meteorological features, and the cloud features extracted in step S5 to form the complete feature vector for that moment. This vector is then extracted using a fixed-time sliding window method to obtain the input feature matrix.

[0108] Furthermore, the meteorological information includes temperature, relative humidity, air pressure, and wind speed;

[0109] Step S7: construct a CNN-LSTM model, input the feature matrix into the convolutional neural network for feature extraction, input the extracted time series features into the long short-term memory layer for irradiance prediction, and obtain the irradiance value for the time period to be predicted. Specifically, step S7 includes the following steps:

[0110] Step S71, constructing a CNN-LSTM model; the CNN-LSTM model includes an input layer, a convolutional layer, a pooling layer, an LSTM layer, a fully connected layer and an output layer;

[0111] Furthermore, the convolution operation is performed in the convolution layer using the following formula:

[0112] (X*W)(i,j)=Σ m Σ n X(i+m,j+n)·W(m,n)

[0113] Where X represents the input data matrix, W represents the convolution kernel (filter), i and j represent the position index of the output matrix, and m and n represent the position index of the convolution kernel.

[0114] Furthermore, the activation function in the convolutional layer is expressed using the following formula:

[0115] a=f(z)

[0116] Wherein, a represents the output value of the convolutional layer, f() represents the activation function of the convolutional layer. In this embodiment, the ReLU activation function is used, and z represents the input value.

[0117] Furthermore, the pooling operation is performed in the pooling layer using the following formula:

[0118] P=max(a)

[0119] Where p represents the output value after pooling, and max() represents the maximum pooling operation.

[0120] like Figure 5As shown in FIG, the CNN-LSTM hybrid neural network provided in this embodiment combines the advantages of convolutional neural networks (CNN) and long short-term memory neural networks (LSTM), and is suitable for processing spatiotemporal data. It has achieved remarkable results in the fields of weather forecasting, image analysis, time series forecasting, etc. Among them, CNN is good at extracting local and spatial features from data, and can capture local patterns in data through convolutional layers and pooling layers. LSTM is a special RNN (Recurrent Neural Network) that can capture long-term dependencies. By adding a gate structure, namely the forget gate f t , input gate i t and output gate o t These three gates effectively address the vanishing and exploding gradient problems in traditional RNNs. Furthermore, by adding cellular memory units to the LSTM network, the network possesses superior memory capabilities, enabling it to address the time series prediction problem of ultra-short-term irradiation in photovoltaic power plants.

[0121] Furthermore, the gate unit information processing is represented by the following formula in the LSTM layer:

[0122] i t =σ(w i,x x t +w i,h h t-1 +b i )

[0123] f t =σ(w f,x x t +w f,h h t-1 +b f )

[0124] o t =σ(w o,x x t +w o,h h t-1 +b o )

[0125] g t =tanh(w g,x x t +w g,h h t-1 +b g )

[0126] Where g t Represents the value of the candidate memory unit, that is, the input x is processed by the activation function t and the previous hidden state h t-1 σ represents the activation function of the LSTM layer. This embodiment uses the sigmoid function.i,x 、w i,h 、w f,x 、w f,h 、w o,x 、w o,h 、w g,x 、w g,h represents the weight matrix, x t represents the input at time t, h t-1 represents the short-term state of the output at time t-1, b i 、b f 、b o 、b g Represents the input gate, forget gate, output gate, candidate, and bias terms of the unit.

[0127] Furthermore, after the gate unit information is processed, the memory module is updated using the following formula:

[0128] c t =f t c t-1 +i t g t

[0129] Where c t represents the long-term state after the improved gradient transfer method, c t-1 Indicates the state of the memory unit at the previous moment.

[0130] Furthermore, since the memory module and the output gate jointly determine the output, this embodiment uses the following formula to calculate the LSTM layer output:

[0131] y t =h t =o t ×tanh(c t )

[0132] Where y t Represents the output value of the LSTM layer.

[0133] Step S712: training the convolutional neural network, specifically, including the following steps:

[0134] I. Prepare training data and split the training data into training set and test set;

[0135] II. Normalize the input signal and input it into the CNN convolutional layer, which selects the convolution kernel size and pooling method.

[0136] III. Initialize all weight parameters in the CNN-LSTM combination model;

[0137] IV. Continuously calculate the neural network results of each layer layer by layer, and finally obtain the output of the neurons in this layer through the activation function;

[0138] V. Use the Adam optimizer to update the weight parameters and bias parameters and backpropagate the training error;

[0139] VI. Based on the training error, calculate the gradient of each parameter, gradually update the model parameters layer by layer, and iterate until the total error converges, then stop training.

[0140] Step S72, inputting the feature matrix into the convolutional neural network for feature extraction;

[0141] Step S73: Input the extracted time series features into the long short-term memory layer for irradiance prediction.

[0142] In this embodiment, the time series feature data is first input from the input layer to the convolution layer, and multiple convolution layers and pooling layers are used to extract the spatial features of the input data. The output of the convolution layer is flattened into a one-dimensional array and passed to the LSTM layer. By using the LSTM layer to capture the temporal features of the data and understand the long-term dependencies, the output of the LSTM layer is passed to the fully connected layer for further nonlinear combination and prediction. Finally, it passes through the output layer to output the prediction result.

[0143] The present invention uses a single-photon lidar to scan the cloud information around the photovoltaic power station, detects key cloud parameters such as cloud coverage, cloud thickness, and cloud height in the area, extracts cloud characteristics, combines historical irradiance and meteorological data, and uses the CNN-LSTM time series prediction algorithm to achieve high-precision irradiance prediction.

[0144] The present invention uses an improved differential zero-crossing algorithm to invert the echo signal profile of the lidar to obtain a cloud parameter data set. Based on the volume scanning mode of the radar, the situation of clouds in the sky can be inverted. The radar scanning range is larger than that of ground-based cloud images taken with a fisheye lens, and the data resolution is higher than that of satellite cloud images. It can also obtain more characteristic parameters of clouds, ensuring the integrity and accuracy of the data.

[0145] The present invention further calculates the sun's path through a solar position model, segments the effective scanning area to extract cloud features, reduces feature dimensions, and finally constructs a feature matrix combining meteorological characteristics and historical irradiation data. The CNN-LSTM hybrid model is then used to predict irradiation, demonstrating high accuracy and stability in forecasts of different weather types.

[0146] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A method for predicting ultra-short-term irradiation in photovoltaic power plants based on single-photon lidar, characterized in that: The following steps are involved: Step S1: Setting the scanning mode of the single-photon laser radar to a volume scanning mode, scanning the sky near the photovoltaic power station at a fixed elevation angle to obtain echo signals in different directions; Step S2: For each echo signal acquired by the laser radar scan, select the echo signal within 50-100 meters of the highest detection target distance, use its mean as the background noise, and subtract the background noise from the original echo signal to obtain the corrected echo signal; Step S3, using the differential zero-crossing method to identify cloud information in each echo signal; Step S4, obtaining scanning cycle data obtained by rotating the laser radar within the azimuth range, calculating the variation range of the solar azimuth angle of the laser radar within one scanning cycle, and intercepting the key area data according to the time point closest to the solar azimuth angle in the scanning cycle to obtain an intercepted data set; Step S5: reducing the data dimension and extracting more accurate cloud features; Step S6: Select meteorological data and historical irradiation data recorded by the meteorological monitoring platform of the photovoltaic power station, delete missing data and erroneous data, obtain historical time-series irradiation and time-series meteorological characteristics, merge the historical time-series irradiation, time-series meteorological characteristics, and the cloud layer characteristics extracted in step S5 into all feature vectors at that moment, and intercept them using a fixed time interval sliding window method to obtain an input feature matrix; Step S7: construct a CNN-LSTM model, input the feature matrix into the convolutional neural network for feature extraction, input the extracted time series features into the long short-term memory layer for irradiance prediction, and obtain the irradiance value for the time period to be predicted.

2. The method for predicting ultra-short-term irradiation of a photovoltaic power station based on single-photon laser radar according to claim 1, characterized in that: The step S1 comprises the following steps: Step S11, setting the elevation angle of the laser radar to a fixed value. At the fixed elevation angle, the laser radar rotates around the current azimuth angle α at a constant rate; Step S12: After the laser radar emits a single-photon laser pulse at the current azimuth angle α, the laser radar receives an echo signal P from the direction. α (r i ), the signal strength received by the single-scan laser radar is Among them, r1, r2, ..., r N Indicates the distance of the detected target, N indicates the total number of echo signal sampling points, and i indicates the i-th echo signal sampling point; In step S13, after completing the signal acquisition at the current azimuth angle α, the laser radar continues to change the azimuth angle, emitting single-photon laser pulses and receiving echo signals to the next azimuth angle α+Δα until the predetermined azimuth angle range is covered.

3. The method for predicting ultra-short-term irradiation of a photovoltaic power station based on single-photon laser radar according to claim 2, characterized in that: The step S3 comprises the following steps: Step S31: distance correction is performed on the original echo signal to eliminate signal attenuation due to distance attenuation. The following formula is used to perform distance correction on the original echo signal: X α (r)=P′ α (r)r 2 Where P′ α (r) represents the backscattered signal of the lidar at distance r, X α (r) represents the value of P′ α (r) The signal after distance square correction, r represents the distance of the detected target; Step S32, differentiating the pre-processed echo signal and fitting the differential signal using a sliding window; Step S33: Determine the cloud base and cloud peak positions in the echo signal profile using the differential zero-crossing method to determine the cloud base signal r b 、Yunfeng Signal m 、Genting Signal t , preliminarily screen out candidate signal areas containing cloud signals; Step S34 , screening cloud areas using a threshold value, eliminating misjudgments caused by thick aerosols or background light based on the signal intensity ratio, and determining the number of clouds.

4. The method for predicting ultra-short-term irradiation of a photovoltaic power station based on single-photon laser radar according to claim 3 is characterized in that: The step S34 uses the following formula to filter the cloud area: Where thr represents the threshold, X α (r m ) represents the cloud peak height, X α (r b ) represents the cloud base height, P′(r m ) represents the distance from the cloud peak signal r m The backscattered signal of the laser radar at b ) represents the distance from the cloud base signal r b The lidar backscatter signal at is judged to contain cloud signals when the threshold is greater than 4, and is judged to contain aerosol signals when the threshold is less than or equal to 4.

5. The method for predicting ultra-short-term irradiation of a photovoltaic power station based on single-photon laser radar according to claim 3, characterized in that: The step S4 comprises the following steps: Step S41, calculating the variation range of the solar azimuth angle of the laser radar within one scanning cycle; Step S42, within a scanning cycle of the laser radar, based on the time point corresponding to the closest solar azimuth angle in the scanning cycle, intercept the scanning data of the adjacent 30 seconds, a total of 1 minute of data.

6. The method for ultra-short-term irradiation prediction of a photovoltaic power station based on single-photon laser radar according to claim 5, characterized in that: The step S41 calculates the variation range of the solar azimuth angle of the laser radar within one scanning cycle using the following formula: H=15°×(T-12) Where δ represents the solar declination angle, n represents the nth day of the year, H represents the solar hour angle, T represents the solar time, φ represents the geographical latitude, and α represents the solar hour angle. sun Indicates the solar azimuth.

7. The method for predicting ultra-short-term irradiation of a photovoltaic power station based on single-photon laser radar according to claim 5, characterized in that: The cloud characteristics in step S5 include solar altitude angle, solar azimuth angle, periodic cloud fraction, cloud base height, cloud thickness, number of cloud layers, and regional cloud fraction.

8. The method for predicting ultra-short-term irradiation of a photovoltaic power station based on single-photon laser radar according to claim 7, characterized in that: The step S5 comprises the following steps: Step S51: analyzing the scan cycle data obtained by rotating the lidar within the azimuth range according to step S3, inverting each lidar echo signal to determine whether there is a cloud signal. If there is a cloud signal, determining the cloud base height, cloud thickness, and cloud number; Step S52: Count the number of echo signals containing cloud signals in the entire scanning cycle data, and compare it with the number of echo signals in the entire scanning cycle to obtain the cloud score in the entire scanning cycle, which is used to represent the cloud coverage ratio in the sky during the entire scanning cycle; Step S53: extracting the cloud base height corresponding to the echo signal in the intercepted data set and calculating the mean of all cloud base heights in the key area; wherein, if there are multiple layers of clouds in the key area, the cloud base height of the lowest layer is taken as the representative value of the cloud base height in the area; Step S54: extracting the mean cloud thickness in the key area; if there are multiple layers of cloud in the key area, the thickness of each layer of cloud is accumulated to obtain the total cloud thickness; Step S55, counting the number of cloud layers obtained by inverting the echo signals in the key area, and analyzing the mode of the number of cloud layers in the area to characterize the number of cloud layers in the area; Step S56 , counting the number of echo signals containing cloud signals in the key area, and comparing it with the number of echo signals in the entire key area, to obtain a regional cloud score in the key area.

9. The method for predicting ultra-short-term irradiation of a photovoltaic power station based on single-photon laser radar according to claim 8, characterized in that: The meteorological information in step S6 includes temperature, relative humidity, air pressure, and wind speed.

10. The method for predicting ultra-short-term irradiation of a photovoltaic power station based on single-photon laser radar according to claim 8, characterized in that: The CNN-LSTM model in step S7 includes an input layer, a convolutional layer, a pooling layer, an LSTM layer, a fully connected layer, and an output layer.

Images