A photovoltaic output combination prediction model generation method and device

Abstract

The present disclosure relates to a photovoltaic output combination prediction model generation method and device. The method comprises: generating meteorological features based on historical meteorological prediction data, extracting effective power generation data based on historical power generation data, and establishing a corresponding relationship between the effective power generation data and the meteorological features; analyzing the time stamp of the meteorological features, and calculating the solar azimuth angle and the solar elevation angle based on the time stamp; taking the solar azimuth angle, the solar elevation angle, and the meteorological features as inputs, respectively establishing a point prediction model and a time series prediction model, and completing the training of the point prediction model and the time series prediction model according to the corresponding relationship between the effective power generation data and the meteorological features; generating a combination prediction model of the point prediction model and the time series prediction model based on the bias-variance weight, and completing the construction of the photovoltaic output combination prediction model. The present disclosure improves the prediction accuracy and the reliability of the power system through the combination prediction model.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the field of new energy, in particular to a photovoltaic output combination prediction model generation method and device. BACKGROUND

[0002] With the increasingly serious environmental problems and carbon emission problems caused by energy shortage and traditional energy, the “double carbon” strategy emerges as the times require, and the development of new energy has become an inevitable demand. In recent years, new energy represented by photovoltaic power generation has developed most rapidly. The traditional power system dominated by thermal power has shifted to a new type of power system dominated by new energy. With the rapid increase in photovoltaic power generation penetration, the inherent randomness, volatility and intermittency of photovoltaic power generation have brought increasingly severe challenges to the stable operation of the power system. Accurate prediction of photovoltaic power generation is an effective means to overcome its randomness and other problems, is crucial for formulating accurate power operation scheduling plans in advance, reducing light abandonment and improving photovoltaic power consumption rate, and is an important guarantee for safe and reliable operation of new type of power system. Due to the randomness of new energy generation, accurate prediction of photovoltaic power generation can reduce the difficulty of formulating operation plans for photovoltaic power stations, improve the utilization efficiency of energy storage, reduce the loss caused by generation uncertainty, and improve the economic benefit of photovoltaic power stations.

[0003] The main methods for current photovoltaic power generation prediction are model-based and data-based methods, among which data-based methods have the advantage of high prediction accuracy, have attracted widespread attention in recent years and have been widely applied, mainly including multilayer perception (MLP), support vector regression (SVR), extreme learning machine, ensemble tree model, recurrent neural network (RNN) and other methods. MLP is a neural network that has been applied in new energy generation prediction for a long time and is relatively mature, which can effectively extract point-to-point mapping information. Recurrent neural networks such as long short-term memory (LSTM) and GRU can extract time series information, and have unique advantages in short-term and ultra-short-term prediction of photovoltaic power generation compared to point-to-point mapping models such as MLP. On the other hand, ensemble tree models such as random forest, XGBoost and LightGBM have excellent modeling performance and are less likely to overfit than neural network models, and are widely used in multiple regression prediction. XGBoost based on boosting trees has been applied in various scientific research problems and prediction algorithm competitions due to its fast running speed and good effect.

[0004] Since different prediction models have different advantages and disadvantages, combining different models can take advantage of their respective strengths, and has higher accuracy and stability than a single prediction model. Studies have shown that by combining multiple regression (GB), quantile random forest (QRF), neural network (NN), and k-NN using error weights, the combined model has better prediction results than a single model. Some scholars have proposed a combination of Prophet and LSTM models based on the least squares method, and have achieved good prediction results. Others have used a dynamic weight allocation method based on variance to combine three different prediction models, achieving better prediction accuracy than a single model.

[0005] The disadvantages of the prior art are as follows:

[0006] Disadvantage 1: The influence of the angle of incidence of the sun on the power generation of fixed photovoltaic panels is not considered. Current data-based prediction methods mostly improve prediction accuracy from the perspective of prediction models, and few consider the mechanism. The principle of a data-based prediction model is to extract as much relevant information as possible from the input and output of the data set. The amount of effective information contained in the data set determines the upper limit of the prediction ability of the model. The current mainstream method for short-term prediction of photovoltaic output based on data is to directly establish an end-to-end mapping between meteorological forecasts and power based on numerical weather prediction (NWP). The radiation received by a tilted fixed photovoltaic panel is related to the intensity of solar radiation and the angle of incidence of sunlight. However, the radiation forecast in numerical weather prediction (NWP) usually only has horizontal radiation flux, which is the total radiation received by a unit horizontal surface, including direct sunlight and air scattering. There is no radiation data for a specific inclined surface. However, the current method directly uses the horizontal solar radiation intensity and other meteorological data to predict power generation, without considering the influence of the angle of the sun on power generation, which will significantly limit the prediction effect.

[0007] Disadvantage 2: Few point prediction and time series prediction methods are used for combined prediction. Current combined prediction methods mostly combine the same type of prediction models, such as combining multiple regression, MLP, k-NN, and tree models for point prediction. Few combined prediction methods use both single prediction models and time series prediction models, such as combining point prediction models with LSTM and GRU time series prediction models, so as to take advantage of both point prediction and time series prediction to improve prediction results.

[0008] Disadvantage 3: The prediction bias and prediction variance are not considered simultaneously when combining models. Current mainstream model combination methods are based on the least squares method and the variance-based method. The combination method based on the least squares method needs to invert all the training set feature combinations, which is computationally intensive or difficult to implement when there are many features or redundant features. The combination method based on variance combines modelsi a reciprocal of the prediction variance Vi 1 / Vi As the output weight of the model, when the prediction error presents a Gaussian distribution, according to the weighted least square estimation, the optimal weight is a reciprocal of the prediction variance, that is, Vi 1 / Vi Since in practice, the prediction error may be affected by non-random factors such as defects of the model itself, overfitting or underfitting in addition to random errors in the training sample, the prediction error itself may not necessarily obey the Gaussian distribution, and its deviation may not be 0, at this time the optimal weight will no longer be 1 / Vi Moreover, this method does not consider the prediction deviation, and only considers the variance for combined weight distribution.

[0009] Therefore, one or more methods are needed to solve the above problems.

[0010] It should be noted that the information disclosed in the above background section is only used to strengthen the understanding of the background of the present disclosure, and therefore can include information that does not constitute prior art known to those of ordinary skill in the art. SUMMARY

[0011] The purpose of the present disclosure is to provide a photovoltaic output combination prediction model generation method, and to at least partially overcome one or more problems caused by the limitations and defects of the related art.

[0012] According to one aspect of the present disclosure, a photovoltaic output combination prediction model generation method is provided, comprising:

[0013] extracting effective meteorological data based on historical meteorological prediction data, and screening the effective meteorological data to generate meteorological features, extracting effective power generation data based on historical power generation data, and establishing a correspondence between the effective power generation data and the meteorological features;

[0014] parsing the timestamp of the meteorological features, and calculating the solar azimuth angle and the solar elevation angle based on the timestamp, respectively;

[0015] establishing a point prediction model and a time series prediction model based on the solar azimuth angle, the solar elevation angle, and the meteorological features as inputs, and training the point prediction model and the time series prediction model according to the correspondence between the effective power generation data and the meteorological features;

[0016] generating a combination prediction model of the point prediction model and the time series prediction model based on the bias-variance weight, and completing the construction of the photovoltaic output combination prediction model.

[0017] In an exemplary embodiment of the present disclosure, the method further comprises:

[0018] ​​Filtering and correcting the historical meteorological prediction data, filling missing data, deleting preset weakly correlated data, and normalizing the historical meteorological prediction data to complete effective meteorological data extraction;

[0019] Filtering and correcting the historical power generation data, filling missing data, deleting preset weakly correlated data, and normalizing the historical power generation data to complete effective power generation data extraction.

[0020] In an exemplary embodiment of the present disclosure, the method further comprises:

[0021] Analyzing meteorological features in the effective meteorological data, the meteorological features including horizontal plane radiation flux, horizontal plane latent heat flux, horizontal plane shortwave radiation flux, horizontal plane longwave radiation flux, cloud cover, air pressure, preset height temperature, preset height humidity, preset height wind speed, and preset height wind direction;

[0022] Establishing Spearman correlation coefficients of the meteorological features and the effective power generation data

[0023] wherein d is the difference between two column ranks, and m is the length of each column;

[0024] Screening the meteorological features based on the Spearman correlation coefficients.

[0025] In an exemplary embodiment of the present disclosure, the method further comprises:

[0026] Calculating solar azimuth angles based on the time stamps ,

[0027] wherein, h, m are the hour and minute of local time, respectively, in 24-hour format, θ t 0° at 12:00, and increasing or decreasing by 15° for each hour increase or decrease.

[0028] In an exemplary embodiment of the present disclosure, the method further comprises:

[0029] Calculating solar elevation angles based on the time stamps

[0030] ,

[0031] wherein, is the latitude of the photovoltaic power station; and δ is the declination, i.e., the latitude of the sun's direct point on earth.

[0032] In an exemplary embodiment of the present disclosure, the method uses the solar azimuth angle, solar altitude angle, and meteorological characteristics as inputs to establish a point prediction model and a time series prediction model, respectively, wherein:

[0033] The point prediction models include MLP, XGBoost and lightGBM point prediction models;

[0034] The time series prediction models include LSTM, GRU, bidirectional LSTM, and convolutional LSTM time series prediction models.

[0035] In an exemplary embodiment of the present disclosure, the method further includes:

[0036] Based on bias-variance weighting Generate a combined prediction model of the point prediction model and the time series prediction model, where is the average forecast deviation, is the prediction variance, p i , q i is the penalty coefficient for error and variance;

[0037] Generate a weighted combined prediction value based on the combined prediction model ,in, y i For the i The actual value of the samples.

[0038] In one aspect of the present disclosure, a photovoltaic output combination prediction method is provided, comprising:

[0039] Collect meteorological data to generate meteorological characteristics, and calculate the solar azimuth and solar altitude angle based on time;

[0040] The solar azimuth angle, solar altitude angle and meteorological characteristics are used as inputs to complete photovoltaic output combination prediction based on the combination prediction model.

[0041] In one aspect of the present disclosure, a photovoltaic output combination prediction model generation device is provided, comprising:

[0042] A meteorological feature generation module is used to extract effective meteorological data based on historical meteorological forecast data, screen the effective meteorological data to generate meteorological features, extract effective power generation data based on historical power generation data, and establish a corresponding relationship between the effective power generation data and the meteorological features;

[0043] A solar angle calculation module, configured to analyze the timestamp of the meteorological feature and calculate the solar azimuth and solar altitude angle based on the timestamp;

[0044] a model training module, configured to establish a point prediction model and a time series prediction model respectively based on the solar azimuth angle, the solar elevation angle, and the meteorological feature, and complete training of the point prediction model and the time series prediction model according to the correspondence between the effective power generation data and the meteorological feature;

[0045] a model construction module, configured to generate a combined prediction model of the point prediction model and the time series prediction model based on a bias-variance weight, and complete construction of the combined prediction model of the photovoltaic power output.

[0046] The photovoltaic power output combined prediction model generation method in the example embodiment of the present disclosure, wherein the method comprises: generating a meteorological feature based on historical meteorological prediction data, extracting effective power generation data based on historical power generation data, and establishing a correspondence between the effective power generation data and the meteorological feature; analyzing a timestamp of the meteorological feature, and calculating a solar azimuth angle and a solar elevation angle based on the timestamp respectively; establishing a point prediction model and a time series prediction model respectively based on the solar azimuth angle, the solar elevation angle, and the meteorological feature, and completing training of the point prediction model and the time series prediction model according to the correspondence between the effective power generation data and the meteorological feature; generating a combined prediction model of the point prediction model and the time series prediction model based on a bias-variance weight, and completing construction of the combined prediction model of the photovoltaic power output. The present disclosure improves the prediction accuracy by the combined prediction model, and improves the reliability of the power system.

[0047] It should be understood that the foregoing general description and the following detailed description are only exemplary and explanatory, and cannot limit the present disclosure. BRIEF DESCRIPTION OF DRAWINGS

[0048] The above and other features and advantages of the present disclosure will become more apparent by describing in detail example embodiments thereof with reference to the attached drawings.

[0049] Figure 1 A flowchart of a photovoltaic power output combined prediction model generation method according to an example embodiment of the present disclosure is shown;

[0050] Figure 2 A solar angle calculation and bias-variance-based photovoltaic power prediction block diagram of a photovoltaic power output combined prediction model generation method according to an example embodiment of the present disclosure is shown;

[0051] Figure 3 A solar azimuth angle θ t and elevation angle θ h Variation in a day

[0052] Figure 4 A schematic block diagram of a photovoltaic output combination prediction model generation device according to an example embodiment of the present disclosure is shown. DETAILED DESCRIPTION

[0053] Example embodiments now will be described more fully hereinafter with reference to the accompanying drawings; however, the example embodiments can be implemented in many different forms and should not be construed as limited to the embodiments set forth herein; rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the concept of the example embodiments to those skilled in the art. Like reference numerals refer to like elements throughout the several views and the description of the figures.

[0054] Moreover, the described features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. In the following description, numerous specific details are provided to give a thorough understanding of embodiments of the disclosure. One skilled in the relevant art will recognize, however, that the

[0055] The block diagrams shown in the accompanying drawings are merely functional entities, and do not necessarily have to correspond to physically independent entities. That is, the functional entities can be implemented in the form of software, or a part of the functional entities or the functional entities can be implemented in one or more software-hardened modules, or the functional entities can be implemented in different network and / or processor devices and / or microcontroller devices.

[0056] In the present example embodiment, a photovoltaic output combination prediction model generation method is first provided; as shown in Figure 1 The photovoltaic output combination prediction model generation method can include the following steps:

[0057] Step S110, effective meteorological data is extracted based on historical meteorological prediction data, and meteorological features are generated by screening the effective meteorological data; effective power generation data is extracted based on historical power generation data, and a corresponding relationship between the effective power generation data and the meteorological features is established;

[0058] Step S120, the time stamp of the meteorological features is parsed, and the solar azimuth angle and the solar elevation angle are respectively calculated based on the time stamp;

[0059] Step S130, the solar azimuth angle, the solar elevation angle, and the meteorological features are taken as inputs to respectively establish a point prediction model and a time series prediction model, and the training of the point prediction model and the time series prediction model is completed according to the corresponding relationship between the effective power generation data and the meteorological features.

[0060] Step S140, generating a combined prediction model of the point prediction model and the time series prediction model based on the bias-variance weight, completing the construction of the photovoltaic power output combined prediction model.

[0061] The photovoltaic power output combined prediction model generation method in the example embodiment of the present disclosure, wherein the method comprises: generating meteorological features based on historical meteorological prediction data, extracting effective power generation data based on historical power generation data, and establishing a correspondence between the effective power generation data and the meteorological features; analyzing the time stamp of the meteorological features, and calculating the solar azimuth angle and the solar elevation angle based on the time stamp respectively; taking the solar azimuth angle, the solar elevation angle, and the meteorological features as inputs, respectively establishing a point prediction model and a time series prediction model, and training the point prediction model and the time series prediction model according to the correspondence between the effective power generation data and the meteorological features; generating a combined prediction model of the point prediction model and the time series prediction model based on the bias-variance weight, and completing the construction of the photovoltaic power output combined prediction model. The present disclosure improves the prediction accuracy and the reliability of the power system by using the combined prediction model.

[0062] In the following, the photovoltaic power output combined prediction model generation method in the example embodiment will be further described.

[0063] Embodiment one:

[0064] Step S110, extracting effective meteorological data based on historical meteorological prediction data, and generating meteorological features by screening the effective meteorological data, extracting effective power generation data based on historical power generation data, and establishing a correspondence between the effective power generation data and the meteorological features.

[0065] In the example embodiment, the method further comprises:

[0066] Filtering and correcting the historical meteorological prediction data, filling in missing data, deleting preset weakly correlated data, and normalizing the historical meteorological prediction data to complete the extraction of effective meteorological data;

[0067] Filtering and correcting the historical power generation data, filling in missing data, deleting preset weakly correlated data, and normalizing the historical power generation data to complete the extraction of effective power generation data.

[0068] In the example embodiment, the method further comprises:

[0069] Analyzing meteorological characteristics in the valid meteorological data, the meteorological characteristics including horizontal radiation flux, horizontal lateral flux, horizontal shortwave radiation flux, horizontal longwave radiation flux, cloud cover, air pressure, temperature at a preset height, humidity at a preset height, wind speed at a preset height, and wind direction at a preset height;

[0070] Establish the Spearman correlation coefficient between the meteorological characteristics and the effective power generation data

[0071] , where d is the difference between the two column ranks and m is the length of each column;

[0072] The meteorological characteristics are screened based on the Spearman correlation coefficient.

[0073] Step S120: parsing the timestamp of the meteorological feature, and calculating the solar azimuth angle and the solar altitude angle based on the timestamp.

[0074] In this exemplary embodiment, the method further includes:

[0075] Calculate the solar azimuth angle based on the timestamp ,

[0076] in, h, m The hour and minute of local time, 24-hour format. θ t At 12:00 it is 0°, and for every hour that passes, it increases or decreases by 15°.

[0077] In this exemplary embodiment, the method further includes:

[0078] Calculate the sun altitude angle based on the timestamp

[0079] ,

[0080] in, is the latitude of the photovoltaic power station; δ is the declination, that is, the latitude of the point on the earth where the sun is directly shining.

[0081] Step S130, using the solar azimuth angle, solar altitude angle, and meteorological characteristics as input, respectively establish a point prediction model and a time series prediction model, and complete the training of the point prediction model and the time series prediction model based on the correspondence between the effective power generation data and the meteorological characteristics.

[0082] In the embodiment of this example, the method uses the solar azimuth angle, solar altitude angle, and meteorological characteristics as inputs to establish a point prediction model and a time series prediction model, respectively, wherein:

[0083] The point prediction models include MLP, XGBoost and lightGBM point prediction models;

[0084] The time series prediction models include LSTM, GRU, bidirectional LSTM, and convolutional LSTM time series prediction models.

[0085] Step S140 , generating a combined prediction model of the point prediction model and the time series prediction model based on the deviation-variance weight, and completing the construction of the photovoltaic output combined prediction model.

[0086] In this exemplary embodiment, the method further includes:

[0087] Based on bias-variance weighting Generate a combined prediction model of the point prediction model and the time series prediction model, where is the average forecast deviation, is the prediction variance, p i , q i is the penalty coefficient for error and variance;

[0088] Generate a weighted combined prediction value based on the combined prediction model ,in, y i For the i The actual value of the samples.

[0089] In this exemplary embodiment, when the present disclosure uses only NWP weather forecasts for photovoltaic power generation prediction, the average prediction accuracy of different models, such as MLP, tree models, and recurrent neural networks, is similar. The solar angle feature can be solved simply based on time and the solar motion model. When used together with meteorological features to train the prediction model, the prediction accuracy of each individual model, such as MLP, tree models, and recurrent neural networks, is significantly improved. A bias-variance-based weighted distribution combination method is used to integrate the prediction results of each point prediction model with the time series prediction model, effectively extracting the mapping relationship between single points and time series and power generation. At different prediction scales, the error and variance of the combined prediction are significantly reduced compared to the individual models, improving the accuracy and stability of predictions under different conditions. By improving the accuracy and stability of predictions, the feasibility of photovoltaic power station operation plans can be effectively improved, the energy storage and backup costs of photovoltaic power stations and the transaction costs between photovoltaic power stations and the power grid can be reduced, and the economic benefits of photovoltaic power stations can be improved. Furthermore, by improving the predictability of photovoltaic power generation, the difficulty of formulating new power system scheduling plans can be reduced, the photovoltaic power generation absorption rate can be increased, and the impact of photovoltaic power generation randomness on the new power system can be reduced, thereby improving the reliability of the new power system.

[0090] Example 2:

[0091] In the embodiment of the present example, reference is made to Figure 2 The solar angle calculation and the bias-variance-based photovoltaic power prediction block diagram, the present disclosure includes the following steps:

[0092] The preprocessing of the historical data set The prediction model based on NWP weather forecast is trained based on historical weather forecast data and historical power generation data, and the data quality directly affects the prediction result. Generally, the historical weather forecast data and the historical power generation data contain a large amount of abnormal data and data in the power limiting stage, so it is necessary to preprocess the data set, including filtering and correcting abnormal data, filling missing data, and deleting data that is difficult to repair or affected by power limiting. Finally, the processed data is normalized, and then the numerical values of various weather characteristics are converted to the same range, so as to improve the speed of model training and the prediction accuracy of the model.

[0093] The weather characteristics are screened Common weather characteristics in NWP weather forecast include horizontal radiation flux, horizontal latent radiation flux, horizontal shortwave radiation flux, horizontal longwave radiation flux, cloud amount, air pressure, 2-meter temperature, 2-meter humidity, 2-meter wind speed, 2-meter wind direction, 10-meter wind speed, 10-meter wind direction, and temperature, humidity, wind speed, wind direction at other heights, etc. The total characteristics can reach more than twenty, and some of them may have weak influence on photovoltaic power generation, which may even cause interference and reduce the prediction accuracy when used for prediction. Therefore, it is necessary to screen the weather characteristics. The relationship between weather characteristics and power generation is a typical nonlinear relationship, and the Spearman correlation coefficient can better reflect the nonlinear correlation between different variables, so the present application uses the Spearman correlation coefficient to analyze the correlation between weather characteristics and photovoltaic power generation. Finally, the weather characteristics with high correlation with photovoltaic power generation are selected for power generation prediction.

[0094] The calculation method of the Spearman correlation coefficient is as follows. Assuming that the data of feature a is X a data column, and the photovoltaic power generation data is X P data column, then the Spearman correlation coefficient between feature a and photovoltaic power generation data X P is

[0095] (1)

[0096] wherein d is the difference between the ranks of the two columns, m is the length of each column. represents complete positive correlation between the two columns, represents complete negative correlation between the two columns, The closer the absolute value is to 1, the higher the correlation between the two columns.

[0097] Modeling the sun's movement and calculating the sun's angle in the sky at any moment in a year can be calculated from the sun's movement model. During the day, the sun's azimuth angle (hour angle) θ t and altitude angle θ h change over time as shown in Figure 3 The azimuth angle θ t is the angle between the projection of the sun's light on the horizontal plane and the due south direction, and the altitude angle θ h is the angle between the sun's light and the horizontal plane, and the angle of the sun can be uniquely determined by θ t and θ h .

[0098] 3.1 Solution of the sun's azimuth angle θ t

[0099] The change of the sun's azimuth angle θ t over time can be calculated by the following formula

[0100] (2)

[0101] where h, m are the hour and minute of local time, respectively, in 24-hour format, θ t 0° at 12:00, and increases or decreases by 15° for each hour increase or decrease, -90° at 6:00, and +90° at 18:00. Therefore θ t can be directly calculated from the local time.

[0102] 3.2 Solution of the sun's altitude angle θ h

[0103] The calculation formula of the sun's altitude angle θ h is as follows:

[0104] (3)

[0105] where is the latitude of the photovoltaic power station; δ is the declination, i.e. the latitude of the sun's direct point on Earth, 0° at the spring and autumn equinoxes, 23°27' at the summer solstice, and -23°27' at the winter solstice. ​​

[0106] Sun declination δ The high-precision calculation method is as follows

[0107] (4)

[0108] (5)

[0109] In the formula, N is the accumulated day, that is, the day in a year. For a specific power station or power station group, the latitude ϕ is a constant value, and it can be proved that when the data model is predicted, the value can use any value because the error Δ ϕ can also be regarded as an implicit parameter of the prediction model, and does not affect the final prediction accuracy. The sun declination δ can be directly calculated from the sampling date. The sun azimuth θ t and the elevation angle θ h will be used as input features for subsequent prediction models.

[0110] Different point prediction and time series prediction models are established. According to the sun elevation angle, the sun azimuth and the screened meteorological features, point prediction models such as MLP, XGBoost and lightGBM and time series prediction models such as LSTM, GRU, bidirectional LSTM (Bi-LSTM) and convolutional LSTM (C-LSTM) are trained, which can effectively extract the single-point and time series mapping relationship between historical data and power generation.

[0111] Model weight distribution based on bias-variance and combined prediction Different prediction models have different advantages and disadvantages, and how to combine them together is the key. The common method is to combine the prediction results of different models by weighting to obtain the final prediction value. The basic principle of weight distribution is to assign higher combination weight to the "better" model. Therefore, the key of the combined model is how to quantitatively evaluate the pros and cons of the model, and then reasonably distribute the weight of each prediction model according to the evaluation index. The present application distributes the combined weight of the model based on the prediction bias and variance as the index.

[0112] Let the average prediction bias of the i-th model be i , and the average prediction variance be , and the average prediction error be

[0113] (6)

[0114] Wherein y k is the actual value of the i-th sample, k k =1... n ;​ the prediction value of the first sample. k

[0115] the prediction variance of the first model is i

[0116] (7)

[0117] the prediction error of the model i E i can be decomposed into bias and variance V i and sample noise σ noise,i , and the following relationship exists between them

[0118] (8)

[0119] assign a bias-variance-based weight to the first prediction model i w i

[0120] (9)

[0121] wherein p i , q i is a penalty coefficient of error and variance, and the smaller the bias and variance, i.e., the higher the prediction accuracy and stability, the higher the corresponding output weight. The sum of the weights of all models is 1. When p i = q i , the method is equivalent to the least square weight fusion method.

[0122] then the weighted combination prediction value is

[0123] (10)

[0124] Compared with the commonly used error-based fusion method, the bias-variance-based fusion method proposed in the present application can balance the prediction accuracy and stability of the output model by adjusting the penalty coefficients of bias and variance p i , q i . For example, if the bias penalty coefficient p is larger, the output weight of the model with larger bias is smaller, and the output weight of the model with smaller bias is larger. Thus, while improving the combination prediction accuracy, the prediction variance can still be controlled.

[0125] ​​​​Further, in the present example embodiment, a photovoltaic output combination prediction method is also provided, the method further comprising:

[0126] Collecting meteorological data to generate meteorological features, calculating a solar azimuth angle and a solar elevation angle based on time;

[0127] Taking the solar azimuth angle, the solar elevation angle, and the meteorological features as inputs, completing photovoltaic output combination prediction based on a combination prediction model.

[0128] It should be noted that although the steps of the method in the present disclosure are described in a specific order in the accompanying drawings, this does not require or imply that the steps must be performed in this specific order, or that all the steps shown must be performed to achieve the desired result. In addition or alternatively, certain steps can be omitted, multiple steps can be combined into one step, and / or one step can be divided into multiple steps, etc.

[0129] Further, in the present example embodiment, a photovoltaic output combination prediction model generation device is also provided. Referring to Figure 4 As shown in the figure, the photovoltaic output combination prediction model generation device 400 can include a meteorological feature generation module 410, a solar angle calculation module 420, a model training module 430, and a model construction module 440. Among them:

[0130] The meteorological feature generation module 410 is configured to extract effective meteorological data based on historical meteorological prediction data, and generate meteorological features by filtering the effective meteorological data; extract effective power generation data based on historical power generation data, and establish a correspondence between the effective power generation data and the meteorological features;

[0131] The solar angle calculation module 420 is configured to analyze the timestamp of the meteorological features, and calculate a solar azimuth angle and a solar elevation angle based on the timestamp, respectively;

[0132] The model training module 430 is configured to take the solar azimuth angle, the solar elevation angle, and the meteorological features as inputs, respectively establish a point prediction model and a time series prediction model, and complete training of the point prediction model and the time series prediction model according to the correspondence between the effective power generation data and the meteorological features;

[0133] The model construction module 440 is configured to generate a combination prediction model of the point prediction model and the time series prediction model based on a bias-variance weight, and complete construction of a photovoltaic output combination prediction model.

[0134] The specific details of each photovoltaic output combination prediction model generation device in the above have been described in detail in the corresponding adaptive rule new energy abnormal data discrimination method, therefore, will not be repeated here.

[0135] It should be noted that, although several modules or units of the photovoltaic power output combination prediction model generation apparatus 400 are mentioned in the above detailed description, such division is not mandatory. Indeed, according to embodiments of the present disclosure, the features and functionalities of two or more modules or units described above can be embodied in one module or unit. Conversely, the features and functionalities of one module or unit described above can be further divided into embodied by multiple modules or units.

[0136] Furthermore, the above-described diagrams are merely schematic illustrations of the processes included in the method according to the exemplary embodiments of the present application, and are not intended for limiting purposes. It is readily understood that the processes shown in the above-described diagrams do not indicate or limit the time sequence of the processes. In addition, it is also readily understood that the processes can be executed synchronously or asynchronously, for example, in multiple modules.

[0137] Other embodiments of the present disclosure will be readily apparent to those skilled in the art upon considering the specification and practice of the present disclosure. The present application is intended to cover any variations, uses, or adaptations of the present disclosure following, in general, the principles of the present disclosure and including such departures from the present disclosure that come within the known and customary practice of the art to which the present disclosure pertains. The specification and examples are to be regarded as exemplary only, and the true scope and spirit of the present disclosure are indicated by the appended claims.

[0138] It should be understood that the present disclosure is not limited to the precise structures herein described and illustrated in the drawings, and that various modifications and changes can be made without departing from the scope thereof. The scope of the present disclosure is indicated by the appended claims.

Claims

1. A photovoltaic output combination prediction model generation method, characterized in that: The method comprises: Extracting effective meteorological data based on historical meteorological forecast data, screening the effective meteorological data to generate meteorological features, extracting effective power generation data based on historical power generation data, and establishing a corresponding relationship between the effective power generation data and the meteorological features; parsing the timestamp of the meteorological feature, and calculating the solar azimuth and solar altitude angle respectively based on the timestamp; Taking the solar azimuth angle, solar altitude angle and meteorological characteristics as input, respectively establishing a point prediction model and a time series prediction model, and completing the training of the point prediction model and the time series prediction model according to the correspondence between the effective power generation data and the meteorological characteristics; Based on bias-variance weighting Generate a combined prediction model of the point prediction model and the time series prediction model, where For the i The average prediction bias of the models, For the i The prediction variance of the model, p i , q i For the i The penalty coefficients of the deviation and variance of each model; generating a weighted combined prediction value based on the combined prediction model ,in, y i For the i The predicted value of each model; based on the combined prediction model of the point prediction model and the time series prediction model, the photovoltaic output combined prediction model is constructed.

2. The method according to claim 1, wherein The method further comprises: Filter and correct the historical meteorological forecast data, fill in missing data, delete preset weakly related data, and normalize the historical meteorological forecast data to complete the extraction of effective meteorological data; The historical power generation data is filtered and corrected, missing data is filled, preset weakly correlated data is deleted, and the historical power generation data is normalized to complete the extraction of effective power generation data.

3. The method according to claim 1, wherein The method further comprises: Analyzing meteorological characteristics in the valid meteorological data, the meteorological characteristics including horizontal radiation flux, horizontal lateral flux, horizontal shortwave radiation flux, horizontal longwave radiation flux, cloud cover, air pressure, temperature at a preset height, humidity at a preset height, wind speed at a preset height, and wind direction at a preset height; The Spearman correlation coefficient between the meteorological characteristics and the effective power generation data was established: , where a is the meteorological characteristic, P is the effective power generation data, and the data of meteorological characteristic a is listed as X a Data column, effective power generation data is X p Data columns, d is the difference between the two column ranks, and m is the length of each column; The meteorological characteristics are screened based on the Spearman correlation coefficient.

4. The method according to claim 1, wherein The method further comprises: Calculate the solar azimuth angle based on the timestamp , in, h、m The hour and minute of local time, 24-hour format. θ t At 12:00 it is 0°, and for every hour that passes, it increases or decreases by 15°.

5. The method according to claim 4, wherein The method further comprises: Calculate the sun altitude angle based on the timestamp : ， in, is the latitude of the photovoltaic power station; δ is the declination, that is, the latitude of the point on the earth where the sun is directly shining.

6. The method according to claim 1, wherein The method uses the solar azimuth angle, solar altitude angle and meteorological characteristics as inputs to establish a point prediction model and a time series prediction model, respectively, wherein: The point prediction models include MLP, XGBoost and lightGBM point prediction models; The time series prediction models include LSTM, GRU, bidirectional LSTM and convolutional LSTM time series prediction models.

7. A photovoltaic output combination prediction method, characterized in that: The method further comprises: Collect meteorological data to generate meteorological characteristics, and calculate the solar azimuth and solar altitude angle based on time; The photovoltaic output combination prediction model is generated based on the photovoltaic output combination prediction model generation method according to any one of claims 1 to 6, with the solar azimuth angle, solar altitude angle and meteorological characteristics as input to complete the photovoltaic output combination prediction.

8. A photovoltaic output combination prediction model generation device, characterized in that: The device comprises: A meteorological feature generation module is used to extract effective meteorological data based on historical meteorological forecast data, screen the effective meteorological data to generate meteorological features, extract effective power generation data based on historical power generation data, and establish a corresponding relationship between the effective power generation data and the meteorological features; A solar angle calculation module, configured to analyze the timestamp of the meteorological feature and calculate the solar azimuth and solar altitude angle based on the timestamp; a model training module for establishing a point prediction model and a time series prediction model respectively using the solar azimuth angle, solar altitude angle and meteorological characteristics as input, and completing the training of the point prediction model and the time series prediction model according to the correspondence between the effective power generation data and the meteorological characteristics; Model building module for bias-variance weighting Generate a combined prediction model of the point prediction model and the time series prediction model, where For the i The average prediction bias of the models, For the i The prediction variance of the model, p i , q i For the i The penalty coefficients of the deviation and variance of each model; generating a weighted combined prediction value based on the combined prediction model ,in, y i For the i The predicted value of each model; based on the combined prediction model of the point prediction model and the time series prediction model, the photovoltaic output combined prediction model is constructed.

Images