A Photovoltaic Power Prediction Method and Device Combining Cloud Map and Adjacent Power Station Cluster

Abstract

The present invention belongs to the technical field of photovoltaic power station power prediction, and discloses a photovoltaic power prediction method and device combining cloud images and adjacent power station clusters. The method includes obtaining non-clear sky data of adjacent power stations for constructing model input data, and establishing a non-clear sky ultra-short-term power prediction model according to the obtained non-clear sky data of the adjacent power stations; training the established non-clear sky ultra-short-term power prediction model to predict the output power of the photovoltaic power station. Through the photovoltaic power prediction method and device combining cloud images and adjacent power station clusters of the present invention, by obtaining the non-clear sky data of adjacent power stations for constructing model input data, and establishing a non-clear sky ultra-short-term power prediction model according to the obtained data; and predicting the output power of the photovoltaic power station by using the trained model, the output prediction of the photovoltaic power station is more accurate.

Description

Technical Field

[0001] The present invention belongs to the technical field of photovoltaic power station power prediction, and particularly relates to a photovoltaic power prediction method and device combining cloud images and adjacent power station clusters. Background Technique

[0002] By making a judgment on the future weather at the location of the photovoltaic power station, it is possible to better predict the future output mode of the photovoltaic power station. If it is possible to judge the moments when there are no cloud clusters blocking the power station in the future, then an accurate prediction of the output mode of the photovoltaic system can be made according to the results in combination with the clear sky model.

[0003] At present, the most advanced atmospheric observation satellite in China is equipped with a multi-channel scanning imaging radiometer, an interferometric atmospheric vertical sounder, a lightning imager, and space environment monitoring instruments, and can obtain cloud images of 14 channels in total. It has produced color satellite cloud images for the first time, generates regional observation images as fast as once a minute, and the highest spatial resolution can reach 500m. The current advanced temporal resolution and spatial resolution provide new ideas for the judgment of future weather conditions in photovoltaic power prediction. According to the different band frequencies of the satellite scanner, satellite cloud images can be divided into two categories: visible light satellite cloud images, water vapor channels, and infrared satellite cloud images. The wavelength range of visible light cloud images is 0.55 - 0.75um, and the wavelength range of infrared cloud images is 10.5 - 12.5um. The general area between the two is called the water vapor channel. Commonly used for corresponding to ground irradiance are mostly visible light cloud images and infrared cloud images.

[0004] Visible light cloud images are obtained by using the scanning radiometer of a meteorological satellite to capture the sunlight reflected by the cloud top and the ground. Therefore, they can only be imaged during the day when there is light shining on the cloud top or the ground. The thickness information of the cloud layer can be obtained from visible light cloud images. Thick cloud layers have strong reflection ability and appear white in visible light cloud images with large gray values; conversely, when the cloud layer is thinner, the cloud image appears dark gray with smaller image gray values.

[0005] Infrared cloud images are obtained by converting the radiation in the infrared band into images by an infrared measuring instrument and can represent the temperature at the corresponding position. When the cloud layer is higher, the temperature of the cloud layer will be relatively low at this time, and it appears bright white on the infrared cloud image; conversely, when the cloud cluster is close to the ground, the temperature of the cloud layer will be higher, and the corresponding position on the infrared cloud image appears dark gray. Therefore, infrared cloud images can judge the temperature of the cloud cluster through gray scale and then judge the height of the cloud top. Since infrared remote sensing does not require the reflection of visible light, infrared cloud images are not limited by the fact that visible light cloud images can only be imaged during the day and can be remotely sensed day and night, providing more information than visible light cloud images. However, due to the limitations of current technology, the resolution of infrared cloud images often tends to be lower than that of visible light cloud images. Taking the Fengyun-4 satellite as an example, only the visible light cloud image resolution reaches 500m, while the highest resolution of the infrared cloud image near 12um is only 4km.

[0006] The information sources for weather judgment used in the method for photovoltaic power prediction are mostly numerical weather prediction or ground-based sky cloud images. Ultra-short-term photovoltaic power prediction requires predicting the output of a photovoltaic power station for the next four hours every 15 minutes. Due to the problems of update time and frequency, numerical weather prediction cannot well adapt to the ultra-short-term four-hour prediction time scale. For ground-based sky cloud images, only a small part of the sky image can be captured, as an estimate of the future situation within the next half hour to one hour, and it also cannot well adapt to the ultra-short-term prediction of photovoltaic power. Summary of the Invention

[0007] The object of the present invention is to provide a photovoltaic power prediction method and device combining cloud images and adjacent power station clusters to solve the problems in the prior art:

[0008] To achieve the above object, the present invention provides the following technical solutions:

[0009] A photovoltaic power prediction method combining cloud images and adjacent power station clusters includes:

[0010] Obtain the non-clear sky data of adjacent power stations for constructing model input data, and establish a non-clear sky ultra-short-term power prediction model according to the obtained non-clear sky data of adjacent power stations; train the established non-clear sky ultra-short-term power prediction model to predict the output power of the photovoltaic power station.

[0011] Further, obtaining the non-clear sky data of adjacent power stations for constructing model input data includes: all power and time delays of adjacent power stations, satellite cloud images, and numerical weather prediction data.

[0012] Further, the construction method of the model input data is:

[0013] Input all power station power and time delays, satellite cloud images, and numerical weather prediction data; according to different time periods, respectively perform feature importance ranking on the historical data of adjacent power stations using the extreme gradient boosting algorithm; determine the adjacent power stations included in the top ten feature quantities in terms of importance in the extreme gradient boosting algorithm to form a set of relevant power stations; combine the historical data of the relevant power stations, the historical power data of the target power station, and the important feature quantities included in the numerical weather prediction to construct model input data with the same number of groups as the number of relevant power stations, and complete the construction of the model input data.

[0014] Further, the training method of the non-clear sky ultra-short-term power prediction model is: for each relevant power station as auxiliary data, respectively construct input data;

[0015] Use the data trained by the long short-term memory neural network to obtain prediction models with the same number as the number of relevant power stations.

[0016] Furthermore, the training method of the model is as follows:

[0017] For each relevant power station as auxiliary data, input data is constructed respectively; historical power data of the photovoltaic power station, numerical weather forecast temperature data, numerical weather forecast irradiance data, and numerical weather forecast heat flux are selected as input features, and future power data of the photovoltaic power station is selected as output data; training data based on the long short-term memory neural network is used to obtain prediction models with the same number as the relevant power stations; the motion direction of the cloud cluster is calculated using the block matching algorithm for adjacent two cloud images, and the included angle between the vector pointing from the relevant power station to the target power station and the motion direction is calculated; the historical data of the relevant power station with the smallest included angle is selected as the prediction model trained with the auxiliary data.

[0018] Furthermore, the important feature quantities included in the numerical weather forecast include:

[0019] All the feature quantities of the numerical weather forecast are input before training the prediction model, and the important feature quantities included in the numerical weather forecast with high correlation selected as the input of the prediction model are used.

[0020] Furthermore, the objective function of the prediction model is:

[0021]

[0022] Where: is the error loss function of the model, is the complexity of the decision tree of the extreme gradient boosting algorithm, where: is the predicted value, y i is the actual value, f k is the decision tree function.

[0023] Furthermore, the is calculated as follows:

[0024]

[0025] Where: Xi represents the i-th sample, yi represents the predicted value of the i-th sample, and the regression tree space F is:

[0026] F = {f(X) = ω q(X)}(q:R m →T, ω∈R T )

[0027] In the formula, q represents the structure of each tree; T represents the number of leaf nodes.

[0028] Further, the method for predicting the output power of a photovoltaic power station is as follows: Based on a long short-term memory neural network, a mapping relationship between multi-data sources and the target future power is established to predict the future power, where: the mapping relationship between multi-data sources and the target future power is:

[0029] f(τ) = σ[W f x(τ)+U f h(τ - 1)+b f

[0030] i(τ) = σ[W i x(τ)+U i h(τ - 1)+b i

[0031]

[0032] o(τ) = σ[W o x(τ)+U o h(τ - 1)+b o

[0033]

[0034]

[0035] In the formula: h(τ - 1), c(τ - 1), and x(τ) are the input quantities of the objective function; h(τ) and c(τ) are the output quantities of the objective function; the rest are the intermediate variables of the objective function.

[0036] A photovoltaic power prediction device combining a cloud map and an adjacent power station cluster includes:

[0037] A data acquisition module for acquiring non-clear sky data of adjacent power stations for constructing model input data;

[0038] A modeling module for establishing a non-clear sky ultra-short-term power prediction model based on the acquired non-clear sky data of adjacent power stations;

[0039] A prediction module for training the established non-clear sky ultra-short-term power prediction model to predict the output power of a photovoltaic power station.

[0040] Compared with the prior art, the advantages of the present invention are:

[0041] The photovoltaic power prediction method of the present invention combining a cloud map and an adjacent power station cluster acquires non-clear sky data of adjacent power stations for constructing model input data, and establishes a non-clear sky ultra-short-term power prediction model based on the acquired data; by predicting the output power of a photovoltaic power station through the trained model, the output prediction of the photovoltaic power station is more accurate. ​​​BRIEF DESCRIPTION OF THE DRAWINGS

[0042] The accompanying drawings forming a part of this invention are used to provide a further understanding of the invention. The schematic embodiments and descriptions thereof of the invention are used to explain the invention and shall not unduly limit the invention. In the drawings:

[0043] Figure 1 is a schematic flow diagram of a photovoltaic power prediction method combining cloud maps and adjacent power station clusters provided by an embodiment of the invention;

[0044] Figure 2 is a schematic diagram comparing the actual photovoltaic output power and the clear sky model predicted power of a photovoltaic power prediction method combining cloud maps and adjacent power station clusters of the invention;

[0045] Figure 3 is a schematic diagram of the decomposition of the actual photovoltaic output power of a photovoltaic power prediction method combining cloud maps and adjacent power station clusters of the invention;

[0046] Figure 4 is the NWP (Numerical Weather Prediction) feature importance ranking of a photovoltaic power prediction method combining cloud maps and adjacent power station clusters of the invention;

[0047] Figure 5 is the correlation between short-wave radiation and power of a photovoltaic power prediction method combining cloud maps and adjacent power station clusters of the invention;

[0048] Figure 6 is the comparison of the NWP short-wave radiation and power curves on a certain day in January and June of a photovoltaic power prediction method combining cloud maps and adjacent power station clusters of the invention;

[0049] Figure 7 is the correlation between the normalized short-wave radiation and power of a photovoltaic power prediction method combining cloud maps and adjacent power station clusters of the invention;

[0050] Figure 8 is the correlation coefficient between the photovoltaic power and temperature of a photovoltaic power prediction method combining cloud maps and adjacent power station clusters of the invention;

[0051] Figure 9 is the correlation coefficient between the photovoltaic power and the heat flux of a photovoltaic power prediction method combining cloud maps and adjacent power station clusters of the invention;

[0052] Figure 10 is the structure of the LSTM memory module of a photovoltaic power prediction method combining cloud maps and adjacent power station clusters of the invention;

[0053] Figure 11Comparison of test results between the embodiment of the present invention and the control group of a photovoltaic power prediction method combining cloud maps and adjacent power station clusters of the present invention;

[0054] Figure 12 Comparison of the predicted values 15 minutes in advance and the measured power curve of each method of a photovoltaic power prediction method combining cloud maps and adjacent power station clusters of the present invention;

[0055] Figure 13 Comparison of the predicted values 4 hours in advance and the measured power curve of each method of a photovoltaic power prediction method combining cloud maps and adjacent power station clusters of the present invention. Detailed implementation manners

[0056] The present invention will be described in detail below with reference to the drawings and in conjunction with embodiments. It should be noted that, without conflict, the embodiments in the present invention and the features in the embodiments can be combined with each other.

[0057] The following detailed descriptions are all exemplary descriptions, aiming to provide further detailed descriptions of the present invention. Unless otherwise specified, all technical terms adopted by the present invention have the same meaning as commonly understood by those of ordinary skill in the technical field to which the present invention belongs. The terms used in the present invention are only for describing specific implementation manners and are not intended to limit the exemplary embodiments according to the present invention.

[0058] Embodiment 1

[0059] Combined with Figure 1 A photovoltaic power prediction method combining cloud maps and adjacent power station clusters according to an embodiment of the present invention is described, including:

[0060] Obtain the non-clear sky data of adjacent power stations for constructing the input data of the model, and establish a non-clear sky ultra-short-term power prediction model according to the obtained non-clear sky data of adjacent power stations; train the established non-clear sky ultra-short-term power prediction model to predict the output power of the photovoltaic power station.

[0061] As Figure 2 And Figure 3 Shown, Figure 2 It is a schematic diagram of the comparison between the actual photovoltaic output power and the clear sky model predicted power of a photovoltaic power prediction method combining cloud maps and adjacent power station clusters of the present invention; from Figure 2 It can be seen that the solid line represents the change trend of the actual power output size as time passes during a day, and the dotted line represents the trend of the power size predicted by the clear sky model. By comparison, it can be obtained that the power change trend predicted by the clear sky model is stable and there are no large fluctuations.

[0062] Figure 3Schematic diagram of the decomposition of the actual photovoltaic output power of a photovoltaic power prediction method combining cloud maps and adjacent power station clusters according to the present invention; Figure 3 In [figure], a is the curve of the output power in clear sky changing with time. It can be seen that the power output in clear sky during power prediction is stable and has no large fluctuations, while Figure 3 In [figure], b is the schematic diagram of the power actually output by the photovoltaic, showing a fluctuating sequence. It can be seen that as time passes during a day, due to reasons such as cloud mass occlusion, the power changes and fluctuates greatly.

[0063] More specifically, a photovoltaic power prediction method combining cloud maps and adjacent power station clusters of the present invention includes the following steps:

[0064] Step 1, construct input data;

[0065] Input the power of all power stations, time delay, satellite cloud maps, and NWP (Numerical Weather Prediction) data;

[0066] According to different time periods, respectively rank the feature importance of the historical data of adjacent power stations using the XGBoost (Extreme Gradient Boosting) algorithm;

[0067] Determine the adjacent power stations included in the top ten feature quantities ranked by importance, and form a set of relevant power stations;

[0068] Combine the historical data of the relevant power stations, the historical power data of the target power station, and the important feature quantities included in the NWP (Numerical Weather Prediction) to construct model input data with the same number of groups as the number of relevant power stations.

[0069] Among them, XGBoost (Extreme Gradient Boosting) is a tree-based learning algorithm that uses decision trees as the basic unit of the model and consists of many decision trees to form a decision forest. The algorithm attempts to correct the residuals of all previous weak learners by adding new weak learners. When these learners are combined for the final prediction, its accuracy rate will be higher than that of a single learner. XGBoost adds the results of K trees as the final prediction value, specifically as follows:

[0070]

[0071] In the formula, Xi represents the i-th sample, yi represents the predicted value of the i-th sample, and the regression tree space F is:

[0072] F = {f(X) = ω q(X)}(q:R m →T, ω∈R T )

[0073] In the formula, q represents the structure of each tree; T represents the number of leaf nodes; the predicted value of XGBoost is the sum of the leaf node values corresponding to each tree. Therefore, we should minimize the following objective function l(Φ) with a regularization term:

[0074]

[0075] In the formula, the first term represents the error loss function of the model, which can be the logistic loss function or MSE (mean squared error), etc.; the second term ∑Ω(fk) is a regularization term, representing the complexity of the tree and preventing overfitting. The regularization term is generally set as follows:

[0076]

[0077] According to the above objective function l(Φ), the model is trained in an additive manner. The objective function in the t-th round of model training is as follows:

[0078]

[0079] Next, perform a Taylor expansion on the objective function, which can be transformed into:

[0080]

[0081] Among them,

[0082] Substitute the optimal value of the leaf node into the objective function, and the final form of the objective function is:

[0083]

[0084] Through the above formula, we can obtain that when the structure of the tree is determined, the structure score of the tree is only related to its first-order derivative and second-order derivative. The smaller the score, the better the structure. In addition, XGBoost uses a greedy algorithm to divide nodes: starting from a single leaf node, iteratively split to add nodes to the tree. The loss function after node splitting:

[0085]

[0086] To prevent overfitting, XGBoost also introduces shrinkage and column subsampling, and uses greedy algorithms and approximate algorithms to find the best split point.

[0087] XGBoost has five common feature importance evaluation methods, which are respectively:

[0088] 1) weight: in the form of weight, indicating how many times a feature is used when splitting nodes in all decision trees;

[0089] 2) gain: (Average) gain form, representing the average gain brought by a feature when splitting nodes in all decision trees;

[0090] 3) cover: (Average) coverage, representing the average number of samples covered when a feature splits nodes in all decision trees;

[0091] 4) total_gain: Relative to gain, representing the total gain size brought;

[0092] 5) total_cover: Relative to cover, representing the total number of samples covered.

[0093] In this embodiment, XGBoost is used to construct a regression model to evaluate the importance of the model input. The input data used (historical power series, numerical weather prediction, etc.) is a time series and are all continuous values. During the construction of the tree, continuous features will generate a very large number of segmentation state spaces. weight evaluates the importance of features by the number of times used when splitting nodes. When using weight as the feature importance calculation method, the robustness of the results of continuous data may not be strong. gain evaluates the importance of features by the information gain of the features to the model. When using gain for feature ranking, the difference between the head and tail values of the sorted order is relatively large. However, since this embodiment is a simple model tuning, gain can point out the most important features.

[0094] Among them, the NWP (numerical weather prediction) in this embodiment includes 22-dimensional features such as temperature, humidity, long and short wave radiation, cloud amount, etc. For the target photovoltaic power prediction in this article, not every dimension of data has a strong correlation with our target. Adding all features as inputs to the prediction model for training, on the one hand, the existence of irrelevant and redundant features may lead to the curse of dimensionality, greatly increasing the amount of data required for model training and increasing the time required for model training; on the other hand, irrelevant and redundant data will also introduce noise into the parameters of the model, affecting the accuracy of the model. Therefore, this embodiment selects a part of important feature quantities from the NWP multi-dimensional data as input data to improve the accuracy of the prediction model.

[0095] Specifically, the important feature quantities of NWP in this embodiment include:

[0096] All feature quantities of NWP are input before training the prediction model; use XGBoost to rank the correlation between each feature of the numerical weather prediction and the photovoltaic power of the target power station, as Figure 4 shown; select features with high correlation as the important feature quantities included in the NWP (numerical weather prediction) input to the prediction model. By Figure 4It can be seen that the characteristic importance of short-wave irradiance, temperature, and heat flux is significantly higher than that of other characteristics. Specific analyses are conducted on these three characteristics respectively:

[0097] 1) Short-wave radiation

[0098] The annual output of the photovoltaic power station varies with the short-wave radiation in the NWP as Figure 5 shown. It can be seen from this scatter plot that there is a certain positive correlation between the output of the photovoltaic power station and the short-wave radiation, and the Pearson correlation coefficient between the two reaches 0.8447. However, there is also a large dispersion between the two in the figure.

[0099] As Figure 6 shown, it shows the comparison of the irradiance and the power station output normalized by the maximum value on a certain day in January and June of a certain power station; select the curves between the short-wave radiation and the power on specific dates for analysis. From the comparison of c and d in Figure 6 , it can be seen that the gap between the solid line and the dashed line in d is significantly reduced, that is, as the difference between the power and the NWP short-wave radiation changes with time, the difference between the two is not large, or there is the same change trend, and within a certain range of changes, for example: the relative amplitude of the short-wave radiation in winter in January is significantly lower than that in summer in June. Due to the influence of the earth's revolution, the irradiance reaching the ground will vary with the seasons, so the short-wave radiation in numerical weather prediction also has similar characteristics. If this data is directly used for model training, it is bound to not be able to well utilize the mapping relationship between this feature and the power in historical data.

[0100] Based on the above analysis, in order to remove the characteristics of the relationship between the NWP short-wave radiation and the power changing with the seasons, this section considers normalizing the seasonal trend term in the NWP short-wave radiation when using the NWP short-wave radiation data. Divide the daily NWP short-wave radiation data by the maximum value of the extraterrestrial irradiance on the corresponding date to obtain the normalized NWP short-wave radiation. The scatter plot between the normalized NWP short-wave radiation and the power is as Figure 7 shown, and the dispersion between the two is greatly reduced, and the Pearson correlation coefficient also increases to about 0.91.

[0101] 2) Short-wave radiation

[0102] The temperature of the photovoltaic system components will affect the efficiency of the photovoltaic system. Therefore, the two characteristics of temperature and heat flux in the NWP prediction values that can reflect the temperature of the photovoltaic system components are also selected as characteristics with relatively strong importance. Calculate the Pearson correlation coefficients between the power and the NWP temperature and irradiance using the annual data of a certain photovoltaic power station in Jilin in 2018, as shown in Table 1, and the daily serial correlation situation is as Figure 8 , Figure 9 shown:

[0103] Table 1 Correlation Coefficients among Photovoltaic Power, Temperature, and Heat Flux

[0104]

[0105] From the perspective of serial correlation coefficients, there is a strong correlation between photovoltaic power and temperature as well as heat flux. From the perspective of daily changes in correlation coefficients, the correlation coefficients between photovoltaic power and the two features of temperature and heat flux fluctuate significantly, but the overall trend fluctuates around a relatively high correlation coefficient.

[0106] In summary, under the existing data conditions, in this embodiment, short-wave radiation, temperature, and heat flux will be selected from NWP as input features of the prediction model, and among them, short-wave radiation will use the extraterrestrial irradiance on a clear sky after removing the seasonal trend term as the input feature.

[0107] Step 2, train the prediction model;

[0108] For each relevant power station as auxiliary data, input data are constructed respectively; the LSTM (Long Short-Term Memory Neural Network) is used to train the data to obtain prediction models with the same number as the relevant power stations.

[0109] Specifically, based on the Long Short-Term Memory Neural Network LSTM, a mapping relationship from multiple data sources to the target future power can be established to predict the future power. The classic LSTM consists of a sequence of input layer, hidden layer, and output layer, and the hidden layer contains some memory units with input and output gates. The improved LSTM introduces a new gate, named the forget gate, into the memory unit, and the information flow in the memory unit is regulated by these gates. The architecture of the memory block is as Figure 10 shown, and the update formula of a memory module at time step τ is as follows:

[0110] f(τ) = σ[W f x(τ)+U f h(τ - 1)+b f

[0111] i(τ) = σ[W i x(τ)+U i h(τ - 1)+b i

[0112]

[0113] o(τ) = σ[W o x(τ)+U o h(τ - 1)+b o

[0114] ​​​

[0115]

[0116] Among them, the special symbol represents element multiplication. In the formula: h(τ-1), c(τ-1), and x(τ) are the input quantities of the objective function; h(τ) and c(τ) are the output quantities of the objective function; the rest are the intermediate variables of the objective function. That is, except for h(t) and c(t) in the formula, the others are intermediate data and have no specific meaning. The above formula will Figure 10 show the data processing process in the architecture of the memory block in

[0117] Step 3, model selection and power prediction.

[0118] Calculate the movement direction of the cloud cluster using two adjacent cloud maps; calculate the angle between the vector pointing from the relevant power station to the target power station and the movement direction; select the historical data of the relevant power station with the smallest angle as the model trained with auxiliary data as the prediction model; predict the photovoltaic output power of the target power station.

[0119] Specifically, to verify the accuracy of the method of the present invention, the photovoltaic power is predicted respectively by the method of only using local historical power data, the method of using the spatio-temporal correlation of cluster power stations based on VAR, the method of predicting NWP data after feature extraction, and the method of predicting NWP and adjacent power stations after feature screening. A control group is established to measure the accuracy of this embodiment. The advantages and disadvantages between the algorithms of each group are measured by the nominal root mean square error (nRMSE). The prediction errors at the moments when the power station has output are statistically counted, as shown in Table 2:

[0120] Table 2 Nominal root mean square error (nRMSE) between the experimental group and the control group

[0121]

[0122]

[0123] The results of the method of the present invention and the control group method on the change of the prediction accuracy in the next 4 hours with the time scale in the test cases are shown in Figure 11: The results of the case using only local historical power data as input are the worst because there is almost no information about future fluctuations in the local historical power data; The control example 2 based on VAR combined with adjacent power plants introduces the information of adjacent power plants, and the prediction accuracy is improved compared with the control example 1 using only local power data; The prediction method of NWP after adding feature screening adds the screened NWP information, providing more information about the future, and the prediction accuracy on the long time scale is greatly improved compared with the previous two methods; On this basis, adding the method of adjacent power plants brings some improvement to the statistical prediction accuracy; The embodiment of the method of the present invention adds the NWP after screening features, and comprehensively selects the model combining the spatio-temporal correlation of adjacent power plants in real time from the satellite cloud images, further improving the prediction accuracy compared with the model only adding NWP.

[0124] Select a period of time to compare the prediction curves of each algorithm 15 minutes and 4 hours in advance with the actual power curve, as shown in Figures 12, Figure 13 shown: From the power curve, the method proposed by the present invention is significantly better than the model using only historical power as input and the prediction algorithm based on VAR combined with adjacent power plants, and the prediction curve is closer to the true value compared with these two models. Comparing the prediction method of NWP after adding feature screening and the prediction method adding adjacent power plant data on this basis, although it cannot be guaranteed that the prediction results at each moment are better than these two models, but from the statistical results, the method of the present invention still has advantages. This is because the relevant power plants judged by the satellite cloud images are not necessarily the most relevant to the target power plant at future moments, but from the perspective of a relatively long time scale, the test results of the method proposed by the present invention are always at a better accuracy level compared with the control models.

[0125] Embodiment 2

[0126] A photovoltaic power prediction device combining cloud images and adjacent power plant clusters, comprising:

[0127] A data acquisition module for acquiring non-clear sky data of adjacent power plants for constructing model input data;

[0128] A modeling module for establishing a non-clear sky ultra-short-term power prediction model according to the acquired non-clear sky data of adjacent power plants;

[0129] A prediction module for training the established non-clear sky ultra-short-term power prediction model to predict the output power of a photovoltaic power plant.

[0130] As is known by common technical knowledge, the present invention can be implemented by other embodiments that do not depart from its spiritual essence or essential features. Therefore, the above-disclosed embodiments are illustrative in all aspects and not exclusive. All changes within the scope of the present invention or within the scope equivalent to the present invention are encompassed by the present invention.

[0131] Those skilled in the art should understand that the embodiments of the present invention can be provided as a method, a system, or a computer program product. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk memory, CD-ROM, optical memory, etc.) that contain computer-usable program code.

[0132] The present invention is described with reference to the flowcharts and / or block diagrams of methods, apparatuses (systems), and computer program products according to the embodiments of the present invention. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, as well as the combination of flows and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, such that the instructions executed by the processor of the computer or other programmable data processing devices generate means for implementing the functions specified in Figure 1 one flow or multiple flows and / or blocks Figure 1 one block or multiple blocks.

[0133] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, such that the instructions stored in the computer-readable memory generate a manufactured article including instruction means that implement the functions specified in Figure 1 one flow or multiple flows and / or blocks Figure 1 one block or multiple blocks.

[0134] These computer program instructions can also be loaded onto a computer or other programmable data processing device, such that a series of operation steps are executed on the computer or other programmable device to generate a computer-implemented process, and thus the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in Figure 1 one flow or multiple flows and / or blocks Figure 1 one block or multiple blocks.

[0135] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Although the present invention has been described in detail with reference to the above embodiments, those of ordinary skill in the art should understand that: modifications or equivalent replacements can still be made to the specific implementation manners of the present invention, and any modifications or equivalent replacements that do not depart from the spirit and scope of the present invention should be covered by the protection scope of the claims of the present invention.

Claims

1. A photovoltaic power prediction method combining cloud maps and adjacent power station clusters, characterized in that, Including: Obtain the non-clear sky data of adjacent power stations for constructing model input data, and establish a non-clear sky ultra-short-term power prediction model based on the obtained non-clear sky data of adjacent power stations; Train the established non-clear sky ultra-short-term power prediction model to predict the output power of the photovoltaic power station. The construction method of the model input data is as follows: Input the power of all power stations, time delay, satellite cloud images, and numerical weather forecast data; According to different time periods, respectively rank the importance of features of the historical data of adjacent power stations using the extreme gradient boosting algorithm. Determine the adjacent power stations included in the top ten feature quantities with the highest importance in the extreme gradient boosting algorithm, and form a set of relevant power stations. Construct model input data with the same number of groups as the number of relevant power stations by combining the historical data of relevant power stations, the historical power data of the target power station, and the important feature quantities included in the numerical weather forecast, thus completing the construction of the model input data. The training method of the model is as follows: For each relevant power station as auxiliary data, construct input data respectively; Select the historical power data of the photovoltaic power station, the temperature data of the numerical weather forecast, the irradiance data of the numerical weather forecast, and the heat flux of the numerical weather forecast as input features, and select the future power data of the photovoltaic power station as output data; Use the long short-term memory neural network to train the data to obtain prediction models with the same number as the number of relevant power stations; Calculate the movement direction of the cloud cluster using the block matching algorithm for two adjacent cloud images, and calculate the angle between the vector pointing from the relevant power station to the target power station and the movement direction; Select the model trained with the historical data of the relevant power station with the smallest angle as the prediction model.

2. The photovoltaic power prediction method combining cloud images and adjacent power station clusters according to claim 1, wherein, The important feature quantities included in the numerical weather forecast include: All feature quantities of the numerical weather forecast input before training the prediction model and the important feature quantities included in the numerical weather forecast selected with high correlation as the input of the prediction model.

3. A photovoltaic power prediction method combining a cloud map and an adjacent power station cluster according to claim 1, characterized in that, The objective function of the prediction model is: Wherein: is the error loss function of the model, is the complexity of the decision tree of the extreme gradient boosting algorithm. In the formula: is the predicted value is the actual value Decision tree function.

4. A photovoltaic power prediction method combining a cloud map and an adjacent power station cluster according to claim 3, characterized in that, The said : Where: Xi represents the i-th sample, yi represents the predicted value of the i-th sample, and the regression tree space F is: In the formula, q represents the structure of each tree; T represents the number of leaf nodes.

5. A photovoltaic power prediction method combining a cloud map and an adjacent power station cluster according to claim 1, characterized in that, The method for predicting the output power of the photovoltaic power station is: Based on the long short-term memory neural network, establish the mapping relationship between multi-data sources and the target future power, and predict the future power, where: The mapping relationship between multi-data sources and the target future power is: wherein: h(τ - 1), c(τ - 1), x(τ) is the input of the objective function; h(τ), c(τ) is the output of the objective function; the rest are intermediate variables of the objective function.

6. A photovoltaic power prediction device combining a cloud map and an adjacent power station cluster, characterized in that, Including: A data acquisition module for obtaining the non-clear sky data of adjacent power stations for constructing model input data; A modeling module for establishing a non-clear sky ultra-short-term power prediction model based on the obtained non-clear sky data of adjacent power stations; A prediction module for training the established non-clear sky ultra-short-term power prediction model to predict the output power of the photovoltaic power station; Among them, the construction method of the model input data is: Input all power station power, time delay, satellite cloud images, and numerical weather forecast data; according to different time periods, respectively rank the importance of features of historical data of adjacent power stations using the extreme gradient boosting algorithm; determine the adjacent power stations included in the top ten feature quantities with the highest importance in the extreme gradient boosting algorithm, and form a set of relevant power stations; combine the historical data of the relevant power stations with the historical power data of the target power station and the important feature quantities included in the numerical weather forecast to construct model input data with the same number of groups as the number of relevant power stations, and complete the construction of the model input data. The training method of the model is as follows: For each relevant power station as auxiliary data, construct input data respectively; select historical power data of photovoltaic power stations, numerical weather forecast temperature data, numerical weather forecast irradiance data, and numerical weather forecast heat flux as input features, and select future power data of photovoltaic power stations as output data; use the long short-term memory neural network to train the data to obtain prediction models with the same number as the relevant power stations; use the block matching algorithm to calculate the movement direction of cloud clusters using adjacent two cloud images, and calculate the angle between the vector pointing from the relevant power station to the target power station and the movement direction; select the model trained with the historical data of the relevant power station with the smallest angle as the prediction model.

Images