Data sample construction method and device, power prediction model training method and device and power prediction method and device

Abstract

The invention discloses a data sample construction method and device, a power prediction model training method and device and a power prediction method and device. The data sample construction method comprises the steps of obtaining first power reference data of a non-clear sky day; obtaining non-clear sky meteorological data of the non-clear sky day and power data of a first historical related day; and performing feature value construction based on the weather type corresponding to the non-clear sky day, the non-clear sky meteorological data, the power data of the first historical related day and the first power reference data to obtain a first non-clear sky sample for training a first initial prediction model. Therefore, by analyzing the regularity of the photovoltaic power generation power changing along with the time and the earth-day astronomical position, the model can identify the influence of the cloud layer on the power generation power, and the power prediction precision is effectively improved.

Description

Technical Field

[0001] The present invention relates to the technical field, and in particular to a data sample construction, power prediction model training, power prediction method and device. Background Art

[0002] Solar power generation has been increasingly widely used due to its advantages of being clean, renewable and easy to distribute. As the proportion of photovoltaic power generation in total power generation increases, high-precision solar photovoltaic power generation forecasting is essential to maintaining the reliability of the power system and maximizing the efficiency of renewable energy consumption.

[0003] In the related art, the photovoltaic power generation prediction method is mainly based on numerical weather forecast. However, in cloudy and rainy weather conditions, the data of numerical weather forecast is unstable, and the accuracy of photovoltaic power generation prediction needs to be improved. Summary of the invention

[0004] The present invention aims to solve at least one of the technical problems in the related art to a certain extent. To this end, the present invention proposes a data sample construction, power prediction model training, power prediction method and device, which enables the model to identify the impact of clouds on photovoltaic power generation by analyzing the regularity of photovoltaic power generation with time and the astronomical position of the earth and the sun, thereby improving the prediction accuracy of the model.

[0005] To achieve the above-mentioned purpose, the first aspect of the present invention proposes a data sample construction method, which includes: obtaining first power benchmark data for a non-clear sky day; wherein the first power benchmark data is obtained based on the power data of a clear sky day adjacent to the non-clear sky day; the first power benchmark data is used to describe the regularity of the photovoltaic power generation power obtained under weather conditions without cloud cover at the geo-sun astronomical position corresponding to the non-clear sky day; obtaining non-clear sky meteorological data of the non-clear sky day, and power data of a first historical related day; wherein the first historical related day is earlier than the non-clear sky day; constructing feature values ​​based on the weather type corresponding to the non-clear sky day, the non-clear sky meteorological data, the power data of the first historical related day and the first power benchmark data to obtain a first non-clear sky sample for training a first initial prediction model.

[0006] To achieve the above-mentioned purpose, the second aspect of the implementation method of the present invention proposes a power prediction model training method, which includes: obtaining non-clear sky power data on a non-clear sky day, and a first non-clear sky sample constructed by a relative data sample construction method; inputting the first non-clear sky sample into a first initial prediction model for prediction to obtain first power prediction data; training the first initial prediction model according to the first power prediction data and the non-clear sky power data to obtain a trained first power prediction model.

[0007] To achieve the above-mentioned purpose, the third aspect of the present invention proposes a power prediction method, which includes: obtaining meteorological forecast data for a day to be predicted, and second power benchmark data for a historical day of the same period; wherein the meteorological forecast data includes the forecast weather type corresponding to the day to be predicted; the forecast weather type includes the forecast whole-day weather type; the date of the historical day of the same period differs from the date of the day to be predicted by an integer number of years; the second power benchmark data is used to describe the regularity of photovoltaic power generation power under weather conditions without cloud cover at the geosun astronomical position corresponding to the day to be predicted; if the forecast whole-day weather type is a non-clear sky type, obtaining power data for a first historical day; wherein the first historical day is earlier than the day to be predicted; constructing first weather characteristic data based on the forecast weather type, the meteorological forecast data, the power data of the first historical day and the second power benchmark data; inputting the first weather characteristic data into the first power prediction model trained by the aforementioned power prediction model training method to obtain the first power prediction result for the day to be predicted.

[0008] To achieve the above-mentioned purpose, the fourth aspect of the present invention proposes a data sample construction device, which includes: a first data acquisition module, used to obtain first power benchmark data of a non-clear sky day; wherein, the first power benchmark data is obtained based on the power data of a clear sky day adjacent to the non-clear sky day; the first power benchmark data is used to describe the regularity of the photovoltaic power generation power obtained under the weather conditions without cloud cover at the geo-sun astronomical position corresponding to the non-clear sky day; a second data acquisition module, used to obtain the non-clear sky meteorological data of the non-clear sky day, and the power data of a first historical related day; wherein, the first historical related day is earlier than the non-clear sky day; a feature construction module, used to construct feature values ​​based on the weather type corresponding to the non-clear sky day, the non-clear sky meteorological data, the power data of the first historical related day and the first power benchmark data, to obtain a first non-clear sky sample for training a first initial prediction model.

[0009] To achieve the above-mentioned purpose, the fifth aspect of the present invention proposes a power prediction model training device, which includes: a sample acquisition module, used to obtain non-clear sky power data on non-clear sky days, and a first non-clear sky sample constructed by the aforementioned data sample construction method; a power prediction module, used to input the first non-clear sky sample into a first initial prediction model for prediction to obtain first power prediction data; a model training module, used to train the first initial prediction model according to the first power prediction data and the non-clear sky power data to obtain a trained first power prediction model.

[0010] To achieve the above-mentioned purpose, the sixth aspect of the present invention proposes a power prediction device, which includes: a third data acquisition module, used to obtain meteorological forecast data for the day to be predicted, and second power reference data for the same historical day; wherein the meteorological forecast data includes the forecast weather type corresponding to the day to be predicted; the forecast weather type includes the forecast whole-day weather type; the date of the same historical day differs from the date of the day to be predicted by an integer year; the second power reference data is used to describe the regularity of photovoltaic power generation under the weather conditions without cloud cover at the astronomical position of the earth and the sun corresponding to the day to be predicted; a historical power acquisition module, used to obtain power data of the first historical day if the forecast whole-day weather type is a non-clear sky type; wherein the first historical day is earlier than the day to be predicted; a feature data construction module, used to construct first weather feature data based on the forecast weather type, the meteorological forecast data, the power data of the first historical day and the second power reference data; a prediction result acquisition module, used to input the first weather feature data into the first power prediction model trained by the aforementioned power prediction model training method to obtain the first power prediction result of the day to be predicted.

[0011] To achieve the above-mentioned purpose, the seventh aspect of the present invention proposes a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the steps of the method described in any of the above-mentioned embodiments are implemented.

[0012] To achieve the above objectives, an eighth aspect of the present invention proposes a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the steps of the method described in any of the aforementioned embodiments are implemented.

[0013] According to the multiple implementation methods provided in this specification, by introducing the feature of power benchmark data, which is used to represent the regularity of changes in photovoltaic power generation power over time and the astronomical position of the earth and the sun in the absence of cloud cover, the model can learn the power change pattern under such ideal conditions (i.e., standard clear skies) and can identify the impact of clouds on photovoltaic power generation power, thereby improving the power prediction accuracy under cloudy, rainy and other weather conditions.

[0014] Additional aspects and advantages of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 A flowchart of a data sample construction method provided according to an implementation of this specification.

[0016] Figure 2 A schematic diagram of a flow chart for constructing a characteristic value according to an implementation manner of this specification.

[0017] Figure 3 A flowchart of another data sample construction method provided according to an embodiment of the present specification.

[0018] Figure 4 The present invention is a flowchart of another data sample construction method provided according to an embodiment of the present specification.

[0019] Figure 5 A flowchart of a power prediction model training method provided according to an embodiment of the present specification.

[0020] Figure 6 A flowchart of another power prediction model training method provided according to an embodiment of the present specification.

[0021] Figure 7a The present invention is a flowchart of another power prediction model training method provided according to an embodiment of the present specification.

[0022] Figure 7b The figure is a flow chart of a daily maximum power prediction algorithm provided according to an embodiment of the present specification.

[0023] Figure 8 The present invention is a flowchart of a power prediction method provided according to an embodiment of the present invention.

[0024] Figure 9a A schematic diagram of a process for obtaining a second power prediction result according to an implementation of this specification.

[0025] Figure 9bThe figure is a flow chart of a power prediction algorithm under sunny day classification provided according to an implementation manner of the present specification.

[0026] Fig.10 A schematic flow chart of another power prediction method provided according to an embodiment of the present specification.

[0027] Fig.11 The present invention is a schematic diagram of a process for correcting a first power prediction result according to an implementation manner of the present specification.

[0028] Fig.12a A flowchart of model training provided according to an embodiment of the present specification.

[0029] Figure 12b The figure is a flow chart of a power prediction algorithm under cloudy or rainy weather classification provided according to an implementation manner of the present specification.

[0030] Fig.12c A flowchart of model reasoning provided according to an embodiment of the present specification. DETAILED DESCRIPTION

[0031] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present invention, and should not be construed as limiting the present invention.

[0032] Given the growing demand for renewable energy and the decreasing cost of building solar photovoltaic stations, photovoltaic power generation is increasingly playing a role in the clean energy acquisition methods of renewable energy. However, this poses a huge challenge to the power system. For example, the influx of large-scale photovoltaic power generation into the grid may cause instability in the power system, that is, there may be an imbalance between power generation and load demand. Therefore, high-precision solar photovoltaic power generation (PV) forecasting is essential to maintain the reliability of the power system and maximize the efficiency of renewable energy consumption.

[0033] In order to meet this challenge, the power prediction method of solar photovoltaic power generation has become an important part of the power system. Among them, the short-term power generation prediction aims to accurately predict the photovoltaic power curve within the next day, providing important reference information for the dispatching decision of the power grid. With the development of meteorological forecasting technology, a large number of photovoltaic power generation prediction methods based on numerical weather prediction (NWP) have emerged. These power prediction methods are mainly based on the laws of the physical mechanism of power generation and data-driven machine learning algorithms, using the historical power generation data and weather data of photovoltaic stations to learn the correlation between photovoltaic power generation and meteorological factors.

[0034] However, although the current short-term power prediction algorithms can generally achieve high prediction accuracy under sunny weather conditions, the prediction accuracy of the power prediction algorithms often drops significantly under cloudy and rainy weather conditions that are more difficult to predict. In addition to the impact of power prediction uncertainty caused by the instability of numerical weather forecast (NWP) data under such weather conditions, the above-mentioned photovoltaic power prediction methods often ignore the laws of solar power generation itself when modeling, that is, solar power generation is affected by solar irradiance, and the change in solar irradiance is affected by the superposition of the astronomical position of the earth and the sun and the near-earth atmospheric clouds. In related technologies, NWP often does not focus on the influence of the earth's solar astronomical laws (that is, the laws of solar irradiation under standard clear sky conditions without cloud cover), resulting in the prediction accuracy of customized power prediction methods for photovoltaic sites not meeting the needs of the power system.

[0035] In order to improve the accuracy of photovoltaic power generation prediction, especially to improve the accuracy of photovoltaic power generation prediction under cloudy and rainy weather conditions, it is necessary to propose a data sample construction, power prediction model training, power prediction method and device. The data sample construction method provided in this specification analyzes the photovoltaic power generation data under historical clear sky weather conditions to obtain clear sky day power benchmark data used to represent the change of photovoltaic power generation power on clear sky days, and then determines the power benchmark data of non-clear sky days adjacent to the clear sky day based on the power benchmark data of the clear sky day. Among them, the power benchmark data of the clear sky day can be used to reflect the change law of photovoltaic power generation under pure clear sky weather conditions without cloud cover (or, called standard clear sky conditions).

[0036] Since in pure clear sky weather conditions, the change in photovoltaic power generation is mainly affected by the change in solar irradiation, and the change in solar irradiation is affected by the change in the astronomical position of the earth and the sun, that is, the change in the position of the earth's rotation and revolution, the power benchmark data of the clear sky day can be used as a manifestation of the earth's astronomical position law or the solar irradiation law of the corresponding date. Therefore, the power benchmark data of the cloudy day can be used to reflect the earth's astronomical position law or the solar irradiation law of the date corresponding to the cloudy day. Based on the weather type, meteorological data and power benchmark data corresponding to the cloudy day, and the power data of the historical day earlier than the cloudy day, the characteristic value of the cloudy day is constructed to obtain the data sample of the corresponding cloudy day. Therefore, the data sample of the cloudy day constructed in the above manner is added with the photovoltaic power generation power law characteristics under the standard clear sky of the corresponding date, that is, the solar irradiation law characteristics (or, the earth's astronomical position characteristics) of the corresponding date are added.

[0037] The power prediction model training method provided in this specification obtains the first power prediction model by training with the data samples of the above-mentioned non-clear sky day, which can enable the model to more accurately learn the basic change laws followed by photovoltaic power generation power under non-clear sky weather conditions such as cloudy or rainy weather conditions, as well as the impact of cloud cover on photovoltaic power generation power, thereby effectively improving the prediction accuracy of the model under non-clear sky weather conditions.

[0038] Through the above-mentioned method of obtaining power benchmark data, the power benchmark data corresponding to each date in a year can be obtained. The power prediction method provided in this specification obtains the power benchmark data of the same historical day in the same period as the date of the day to be predicted in the historical year, and uses the power benchmark data as the power benchmark data of the day to be predicted. If the forecast weather type for the whole day of the day to be predicted is a non-clear sky type, then based on the forecast weather type of the day to be predicted, the meteorological forecast data, the power benchmark data and the power data of the historical day earlier than the day to be predicted, feature construction is performed to obtain weather feature data, which is used to input the first power prediction model for power prediction to obtain the power prediction result of the day to be predicted. Therefore, by proposing a power prediction method based on power benchmark data, the power prediction accuracy of photovoltaic power generation in non-clear sky weather such as cloudy and rainy weather can be effectively improved.

[0039] In some achievable ways, for clear sky days, the data sample construction method provided in this specification constructs characteristic values ​​of clear sky days based on the weather type of clear sky days, meteorological data and power data of historical days earlier than clear sky days, and obtains data samples of corresponding clear sky days. By using the data samples and power data of clear sky days for model training, a second power prediction model that can be used to predict the power of clear sky days is obtained. Furthermore, the data samples of clear sky days and the maximum power data of corresponding clear sky days can be used for model training, so that the trained second power prediction model can be used to predict the maximum power of the day to be predicted when the weather type of the day to be predicted is a clear sky type. According to the maximum power of the day to be predicted and the corresponding power benchmark data, the power prediction result of the day to be predicted can be determined. Therefore, the power of the day to be predicted under the clear sky state is predicted based on the power benchmark data extracted from historical data and the second power prediction model, which can effectively improve the prediction accuracy of photovoltaic power generation in sunny weather.

[0040] This specification provides a method for constructing a data sample. Figure 1 As shown, the data sample construction method may include the following steps.

[0041] S110, obtaining first power reference data on a cloudy day.

[0042] Among them, the first power benchmark data is obtained based on the power data of a clear sky day adjacent to a non-clear sky day; the first power benchmark data is used to describe the regularity of the photovoltaic power generation power obtained under weather conditions without cloud cover at the astronomical position of the earth and the sun corresponding to the non-clear sky day.

[0043] S120, obtaining cloudy weather data of cloudy days and power data of the first relevant historical day.

[0044] Among them, the first historical relevant day is earlier than the non-clear sky day.

[0045] S130, constructing feature values ​​based on the weather type corresponding to the non-clear sky day, the non-clear sky meteorological data, the power data of the first historical related day and the first power benchmark data to obtain a first non-clear sky sample for training a first initial prediction model.

[0046] Among them, the regularity of photovoltaic power generation under weather conditions without cloud cover refers to the change or change characteristics of photovoltaic power generation with time and astronomical position of the earth and the sun under pure clear sky (or standard clear sky) conditions without cloud cover.

[0047] The power benchmark data can reflect the power variation characteristics of photovoltaic power generation under pure clear sky conditions, and can be used as the basic variation law that photovoltaic power generation will follow under any weather conditions. The power benchmark data can be in the form of a power curve, and the power benchmark data is also referred to as a power benchmark curve or a clear sky standard curve in this specification. The first power benchmark data is obtained by inference.

[0048] The Earth-Sun astronomical position refers to the relative position between the Earth and the Sun, including the Earth's rotation position and the Earth's revolution position around the Sun, which can determine the intensity and angle of the Sun's radiation.

[0049] A cloudy day refers to a day when the overall weather is cloudy, rainy, foggy, or other non-sunny weather. The weather type corresponding to a cloudy day may include the overall weather type of the corresponding cloudy day, or the weather type of any period of the corresponding cloudy day.

[0050] The cloudy weather data refers to the actual meteorological data collected for the corresponding cloudy days, which may include field information such as horizontal total radiation, ambient temperature, ambient humidity, wind speed, wind direction, air pressure, and weather type.

[0051] The first historical related day is a historical date of a cloudy day. The power data of the first historical related day has a certain correlation and reference role in predicting the power data of cloudy days. The power data of the first historical related day can be used to provide time series characteristics for the data of cloudy days.

[0052] The first initial prediction model can be used to predict the photovoltaic power generation power on a non-clear sky day.

[0053] In some cases, photovoltaic power generation is affected by solar irradiance data. In non-clear sky weather conditions such as cloudy and rainy weather, if the solar irradiance data measured by a solar irradiator or other equipment is used as a feature to make the model directly learn the relationship between solar irradiance data and power data, the relationship between the measured solar irradiance data and power data may have a large error due to factors such as cloud cover and ambient temperature, thereby reducing the accuracy of the model. Since the influence of cloud cover on power can be ignored under pure clear sky conditions, it is believed that photovoltaic power generation is mainly affected by solar irradiance. Therefore, the change law of photovoltaic power generation under pure clear sky conditions (i.e., power benchmark data) can be used to describe the corresponding change law of solar irradiance. Since the change of solar irradiance under pure clear sky conditions is affected by the change of the astronomical position of the earth and the sun, the power benchmark data can be further used to reflect the astronomical position of the earth and the sun corresponding to the corresponding date.

[0054] Specifically, for a historical cloudy day with known power data and meteorological data, based on the power data of the clear sky days adjacent to the cloudy day, the first power benchmark data of the cloudy day can be inferred or estimated, which is used to describe the regularity that the photovoltaic power generation power should follow if it is clear sky weather (i.e., no cloud obstruction) under the astronomical position of the earth and the sun corresponding to the cloudy day.

[0055] The measured non-clear sky meteorological data of the non-clear sky day and the measured power data of the first historical related day are obtained. A feature vector is constructed according to the weather type of the non-clear sky day, the non-clear sky meteorological data, the first power reference data and the power data of the first historical related day, and a first non-clear sky sample is obtained as a data sample for training a first initial prediction model for predicting photovoltaic power generation on non-clear sky days.

[0056] It should be noted that, for a designated cloudy day, the first historical relevant day can be a relevant day determined through data analysis and other means that has a greater correlation and reference role in predicting the power of the cloudy day, for example, the past day or the past three days of the cloudy day, or it can be several days with similar overall weather types, similar meteorological data, and similar dates to the cloudy day.

[0057] In some embodiments, the first power reference data may be obtained by performing an average calculation on power data of a clear sky day adjacent to a non-clear sky day.

[0058] The non-clear sky meteorological data corresponds to a collection time point (or collection time period), and the weather type of a non-clear sky day may include the day weather type corresponding to the non-clear sky day, and the time point weather type corresponding to the collection time point (or collection time period).

[0059] Exemplarily, for a specified cloudy day, the weather type is cloudy. The sunny days adjacent to the cloudy day include sunny day 1 earlier than the cloudy day, and sunny day 2 later than the cloudy day. According to the measured power data of sunny day 1 and sunny day 2, the average is calculated according to the interval between sunny day 1 and sunny day 2 to obtain the clear sky standard curve of the cloudy day as the first power reference data.

[0060] According to the cloudy weather type and the collected meteorological data such as irradiance, temperature, humidity, wind speed, etc. of the non-clear sky day, as well as the time series characteristics constructed based on the power data of the first historical related day, such as the power value of the past day of the non-clear sky day and the daily maximum and daytime average power values, as well as the clear sky standard curve characteristics of the non-clear sky day, characteristic values ​​are constructed to obtain the data sample of the non-clear sky day.

[0061] For example, for the above-mentioned non-clear sky day, the instantaneous value of meteorological data is collected at a frequency of 15 minutes. From 00:15 to 00:00 the next day, the meteorological data of 96 collection time points of the day are uniformly sampled at a sampling interval of 15 minutes. The meteorological data of each time point corresponds to a weather type at that time point. For example, the weather type corresponding to the point at 01:00 is cloudy, the weather type corresponding to the point at 02:00 is rainy, the weather type corresponding to the point at 04:00 is cloudy, etc.

[0062] After obtaining the first power benchmark data of the clear sky day, characteristic values ​​can be constructed according to the weather type, meteorological data, power data of the first historical related day, and the first power benchmark data of the cloudy sky day corresponding to each collection time point to obtain the data sample of the cloudy sky day.

[0063] In other embodiments, the first power reference data may be obtained by interpolating power data of a clear sky day adjacent to a cloudy sky day.

[0064] Exemplarily, for a specified cloudy day, the sunny days adjacent to the cloudy day include sunny day 3 earlier than the cloudy day, and sunny day 4 later than the cloudy day. Linear interpolation calculation is performed based on the measured power data of sunny day 3 and sunny day 4 to obtain the first power reference data of the cloudy day. The description of the construction of the data sample of the cloudy day is similar to the description of the construction of the above-mentioned data sample, and will not be described in detail.

[0065] It can be understood that, for a specified cloudy day, the clear days adjacent to the cloudy day may include multiple clear days earlier than the cloudy day, and multiple clear days later than the cloudy day. The description of the method for obtaining the first power reference data in this case is similar to the description of obtaining the first power reference data by averaging or interpolation calculation, and will not be described in detail.

[0066] It should be noted that the non-clear sky meteorological data on a non-clear sky day can be collected by equipment such as environmental monitoring instruments and solar irradiation instruments, and is not specifically limited in this specification.

[0067] The method of constructing characteristic values ​​based on the weather type corresponding to a non-clear sky day, non-clear sky meteorological data, power data of the first historical related day and the first power benchmark data can be determined according to actual application requirements, etc., and is not specifically limited in this specification.

[0068] In the above implementation, the power benchmark data is introduced as a feature to represent the regularity of photovoltaic power generation power changes with time and the astronomical position of the earth and the sun in the absence of cloud cover, so that the model can learn the power change law under this ideal condition (i.e., standard clear sky) as the basic law that the photovoltaic power generation system will follow under any weather conditions.

[0069] By introducing power benchmark data for data sample construction, the model can identify the impact of clouds on photovoltaic power generation by comparing the actual power data on non-clear days with the power benchmark data. At the same time, power benchmark data can provide a standardized way to view the changing trend of power generation, which can maintain a certain degree of comparability even under different weather conditions. This helps the model learn more generally applicable rules, rather than just adapting to specific types of weather patterns.

[0070] Therefore, for non-clear sky weather such as cloudy and rainy weather, which is difficult to predict, by extracting the corresponding power change benchmark law under clear sky conditions (solar radiation change benchmark law) as the corresponding clear sky law, and adding the clear sky law to the model learning, a data sample construction method based on the clear sky standard curve is proposed. This can enable the model to simultaneously learn the basic change law of photovoltaic power generation power (that is, the morphological characteristics of the power benchmark data) and the specific impact of clouds on power, and further propose a power prediction method based on the clear sky standard curve, so as to improve the power prediction accuracy under cloudy, rainy and other weather conditions.

[0071] In some embodiments, obtaining the first power reference data based on power data of a clear sky day adjacent to a cloudy sky day includes: performing interpolation calculation based on power data of a first clear sky day adjacent to the cloudy sky day and power data of a second clear sky day to obtain the first power reference data, wherein the first clear sky day is earlier than the cloudy sky day, and the second clear sky day is later than the cloudy sky day.

[0072] The first clear sky day is the clear sky day closest to the clear sky day before the clear sky day, and the second clear sky day is the clear sky day closest to the clear sky day after the clear sky day. The power data of the first clear sky day and the second clear sky day are both known measured data.

[0073] The acquisition timestamps of the power data of the first clear sky day and the power data of the second clear sky day correspond one to one.

[0074] Specifically, for a specified cloudy day, the first cloudy day that is adjacent to and earlier than the cloudy day, and the second cloudy day that is adjacent to and later than the cloudy day can be determined. The first power reference data of the cloudy day can be obtained by performing interpolation calculation based on the collected power data of the first cloudy day and the power data of the second cloudy day.

[0075] Exemplarily, the historical photovoltaic power generation data collected by the photovoltaic station on the historical date is obtained, and the historical date includes cloudy days and sunny days. The daily power data is collected at a frequency of 15 minutes for instantaneous values, that is, from 00:15 to 00:00 the next day, the power generation data at 96 o'clock on the day is obtained by uniform sampling at a sampling interval of 15 minutes.

[0076] For the 96-point power curve of each day in the above historical power data, the power curve of the clear sky day can be filtered out according to the corresponding weather type of each day to construct a clear sky sample subset, and the corresponding date information can be saved. The power curve of the clear sky day is normalized, that is, the 96-point power value of each day can be divided by the maximum power of the day with the day as the smallest unit, so as to process the daily standard curve value into a value between 0 and 1, thereby obtaining the clear sky standard curve of each clear sky day (that is, the power benchmark data of the clear sky day) and storing it in the clear sky sample subset. Through visual analysis, it is found that the power curve of the clear sky day has the characteristics of periodic changes, which is consistent with the cognition of the relative position of the earth's orbit and the sun, and reflects the change law of photovoltaic power generation power under pure clear sky conditions without cloud cover on the corresponding clear sky day.

[0077] For the specified cloudy day in the above historical dates, the date information can be used to determine the first cloudy day before and the second cloudy day after the cloudy day, and the normalized power curve of the first cloudy day and the normalized power curve of the second cloudy day are determined in the sunny day sample subset according to the corresponding date information. Linear interpolation is performed on the normalized power curves of the first cloudy day and the second cloudy day to obtain the clear sky standard curve (i.e., normalized power curve) of the cloudy day as the first power reference data. The specific interpolation process is as follows:

[0078] For a certain date D (i.e., a non-clear sky day D), any point of its 96-point clear sky standard curve is denoted as i (i = 0, 1, 2, ..., 95), and the value p of any point i of its clear sky standard curve is D,i The value p of the corresponding point i in the clear sky standard curve of the previous adjacent clear sky day Da is D-a,i , and the value p of the corresponding point i in the clear sky standard curve of the next adjacent clear sky day D+b D+b,i It is calculated according to the date interval a and b. The calculation formula is as follows:

[0079]

[0080] Thus, according to the values ​​of the 96 points respectively calculated, the clear sky standard curve of the cloudy sky day D is obtained, so as to obtain the power reference data of the cloudy sky day D.

[0081] Further, by way of example, after obtaining the clear sky standard curve of non-clear sky day D according to the above method, the clear sky standard curve of non-clear sky day D and its corresponding date can be stored. When calculating the power reference data of the next non-clear sky day D', assuming that non-clear sky day D' is adjacent to non-clear sky day D and clear sky day D+b respectively, and D-a<D<D'<D+b, then clear sky day D-a is still the clear sky day adjacent to non-clear sky day D' before, and clear sky day D+b is still the clear sky day adjacent to non-clear sky day D' after. At this time, interpolation calculation can still be performed according to the normalized power curve of clear sky day D-a, the normalized power curve of clear sky day D+b, the time interval between non-clear sky day D' and clear sky day D-a, and the time interval between non-clear sky day D' and clear sky day D+b to obtain the power reference data of non-clear sky day D'.

[0082] Alternatively, interpolation calculation can also be performed according to the normalized power curve of non-clear sky day D, the normalized power curve of clear sky day D+b, the time interval between non-clear sky day D' and non-clear sky day D, and the time interval between non-clear sky day D' and clear sky day D+b to obtain the power reference data of non-clear sky day D'.

[0083] It should be noted that other interpolation methods can also be used for interpolation calculation according to actual application requirements, which are not specifically limited in this specification.

[0084] In some embodiments, the non-clear sky day includes multiple collection time points; the weather type corresponding to the non-clear sky day includes the weather type corresponding to any collection time point; the non-clear sky meteorological data includes the first irradiance data and the first air temperature data corresponding to any collection time point. Referring Figure 2 as shown, based on the weather type corresponding to the non-clear sky day, the non-clear sky meteorological data, the power data of the first historical related day, and the first power reference data, eigenvalue construction can include the following steps.

[0085] S210. Obtain the preset working temperature data of the photovoltaic power generation module and the first working temperature data at any collection time point.

[0086] S220. Based on the preset working temperature, the first irradiance data, the first air temperature data, and the first working temperature data corresponding to any collection time point, determine the first harmonic irradiance data corresponding to any collection time point.

[0087] S230. According to the weather type corresponding to each of the multiple collection time points, the non-clear sky meteorological data, the power data of the first historical related day, the first power reference data, and the first harmonic irradiance data corresponding to each of the multiple collection time points, perform eigenvalue construction to obtain the first non-clear sky sample.

[0088] The collection time point can be determined according to the sampling frequency. The preset operating temperature data is the marked operating temperature set by the photovoltaic power generation component. The first operating temperature data is the actual operating temperature of the photovoltaic power generation component when generating electricity at any collection time point.

[0089] Non-clear weather meteorological data includes meteorological data corresponding to any collection time point, which may include irradiance, temperature, humidity, wind speed, etc.

[0090] The multiple collection time points of the power data of the first historical related day correspond one-to-one with the multiple collection time points included in the cloudy day. Similarly, the multiple time points of the first power reference data correspond one-to-one with the multiple collection time points included in the cloudy day.

[0091] The first harmonized irradiance data is the irradiance data after harmonization or correction of the actual solar irradiance data, and can be used to describe the characteristics of the photoelectric conversion efficiency at the corresponding non-clear sky day collection time point.

[0092] In some cases, the temperature of photovoltaic modules will rise due to the actual solar irradiance and air temperature, which will lead to a decrease in photoelectric conversion efficiency. In order to further improve the accuracy of model training, it is necessary to consider the impact of solar irradiance and air temperature on the increase of photovoltaic module temperature and the decrease of photoelectric conversion efficiency when constructing data samples.

[0093] Specifically, for a designated collection time point on a non-clear sky day, the weather type, first irradiance data and first temperature data corresponding to the collection time point, as well as the working temperature data of the photovoltaic power generation component when working at the collection time point can be obtained. Based on the preset working temperature of the photovoltaic power generation component, the first irradiance data and first temperature data corresponding to the designated collection time point, and the working temperature data of the photovoltaic power generation component at the designated collection time point, calculation is performed to obtain the first harmonized irradiance data corresponding to the designated collection time point.

[0094] In combination with the weather type, meteorological data, first harmonic irradiation data and power data of the first historical related day at the corresponding collection time point corresponding to the designated collection time point, as well as the benchmark data of the first power benchmark data at the corresponding collection time point, the eigenvalue of the designated collection time point is constructed, and the eigenvector corresponding to the designated collection time point can be obtained as the sample of the designated collection time point. In this way, eigenvalues ​​are constructed for multiple collection time points respectively, and the first non-clear sky sample can be obtained. That is, according to the weather type, non-clear sky meteorological data, power data of the first historical related day, first power benchmark data and the first harmonic irradiation data corresponding to each of the multiple collection time points, eigenvalues ​​are constructed to obtain the first non-clear sky sample.

[0095] In some embodiments, the harmonized irradiance data may be calculated in conjunction with a physical model of photovoltaic conversion efficiency.

[0096] Exemplarily, historical meteorological data of historical dates collected by photovoltaic stations are obtained, and historical dates include cloudy days and clear days. Among them, the daily measured meteorological data is collected at a frequency of 15 minutes for instantaneous values, that is, from 00:15 to 00:00 the next day, 96 meteorological data of the day are obtained by uniform sampling at a sampling interval of 15 minutes (corresponding to point numbers 0 to 95). The measured meteorological data can be collected by environmental monitors and solar irradiators, including field information such as horizontal total irradiation, ambient temperature, ambient humidity, wind speed, wind direction, and air pressure. The weather type corresponding to each of the 96 sampling time points each day can also be collected from relevant meteorological data sources.

[0097] For 96 sampling time points on a non-clear sky day D, the corresponding operating temperatures of the photovoltaic power generation components at these 96 points can be collected. For any of the 96 sampling time points, the corresponding harmonic irradiance is calculated in combination with the photovoltaic conversion efficiency physical model, where the photovoltaic conversion efficiency physical model considers the irradiance R and the temperature T causing the photovoltaic component temperature T C The increase in the photoelectric conversion efficiency η leads to a decrease in the effect of harmonizing the irradiance characteristics (i.e., reconciled irradiation data) is calculated as follows:

[0098]

[0099] η=1-β(T C -25℃)

[0100]

[0101] Among them, β is the parameter of the control efficiency reduction range, which can be set according to actual needs or actual application conditions; T std The nominal operating temperature set for the photovoltaic panel; W / m 2 It is the unit of solar irradiance.

[0102] According to the above formula, the harmonic irradiance characteristics of 96 points of cloudy day D can be calculated respectively. The past day of cloudy day D (i.e., the day before cloudy day D-1) is taken as the first historical related day, and the first historical related day D-1 corresponds to 96 points of power data. According to the weather type and meteorological data corresponding to the 96 sampling time points of cloudy day D, the power data of the 96 sampling time points of the first historical related day D-1, the power benchmark data of the 96 points of cloudy day D, and the harmonic irradiance data corresponding to the 96 points, the characteristics of the 96 points of cloudy day D are constructed respectively. Among them, the features corresponding to each point include: the weather information corresponding to the point, such as irradiance, temperature, humidity, wind speed, weather type, etc.; time series features constructed based on historical power data (power data of the first historical related day), such as the power lag value of the past day at the same time point (i.e., the power value at the same time point on the first historical related day) and the daily maximum power value of the past day, the average power value during the day (the average power value during the period when the power generation value is greater than zero), etc.; other features include the harmonized irradiance of the point after harmonization using the above-mentioned conversion efficiency model, and the clear sky standard curve features of the point. The clear sky standard curve features corresponding to each point are the power curve values ​​of the normalized power curve (i.e., the first power reference data) extracted from the non-clear sky day D at the corresponding point.

[0103] Furthermore, considering that the time of the day is periodic, that is, 0 o'clock on the current day is 24 o'clock on the previous day, in order to avoid the influence of the point number on the periodicity of time, a timestamp feature can also be added when constructing the feature. That is, the feature corresponding to each point mentioned above also includes a timestamp feature, such as the point number (0-95) of the current point on the current day and the corresponding hour (for example, the hour corresponding to 3 o'clock in the afternoon is 15), and the coded information after the corresponding point number and hour are transformed by the sine function using the sine function.

[0104] In the above implementation, by introducing the characteristics of the harmonized irradiation data when constructing the data sample, the model can more accurately handle the impact of changes in air temperature, changes in the operating temperature of photovoltaic power generation components, etc. on power generation efficiency, thereby improving the accuracy and reliability of the prediction.

[0105] In some embodiments, reference Figure 3 As shown, the data sample construction method may further include the following steps.

[0106] S310, obtaining clear sky meteorological data of the third clear sky day and power data of the second relevant historical day.

[0107] Among them, the weather type corresponding to the third clear sky day is clear sky type; the second historical related day is earlier than the clear sky day.

[0108] S320, constructing feature values ​​based on the clear sky type, clear sky meteorological data and power data of the second relevant historical day to obtain clear sky samples for training the second initial prediction model.

[0109] The clear sky type can be understood as a standard clear day where the influence of cloud cover can be ignored. The clear sky meteorological data is the actual meteorological data collected for the corresponding clear sky day, which can include field information such as horizontal total irradiance, ambient temperature, ambient humidity, wind speed, wind direction, air pressure, and weather type.

[0110] The second historical related day is the historical date of a clear sky day, and the power data of the second historical related day has a certain correlation and reference role in predicting the power data of the clear sky day.

[0111] The second initial prediction model can be used to predict the photovoltaic power generation power on a clear sky day.

[0112] Specifically, for a certain historical clear sky day with known power data and meteorological data, a second historical related day can be determined, and the measured power data of the second historical related day can be obtained. A feature vector is constructed according to the weather type of the clear sky day (i.e., the corresponding clear sky type), the clear sky meteorological data, and the power data of the second historical related day, and a clear sky sample of the clear sky day is obtained as a data sample for training a second initial prediction model for predicting photovoltaic power generation on a clear sky day.

[0113] In some embodiments, for a specified clear sky day, the three days before the clear sky day may be selected as the second historical related day. The feature value may be constructed based on the collected meteorological data of the clear sky day and the power data of the second historical related day.

[0114] Exemplarily, historical meteorological data collected by the photovoltaic station on historical dates are obtained, and the historical dates include cloudy days and clear days. The daily measured meteorological data is collected at a frequency of 15 minutes for instantaneous values, that is, from 00:15 to 00:00 the next day, the meteorological data of 96 points on the day are obtained by uniform sampling at a sampling interval of 15 minutes.

[0115] For a specified clear sky day S, the past three days of the clear sky day are taken as the second historical related days, that is, S-1, S-2, and S-3 are taken as the second historical related days. Each day of the second historical related day corresponds to the above-mentioned corresponding 96 points of power data. According to the weather type of sunny day, 96 points of meteorological data of clear sky day S, 96 points of power data of S-1, 96 points of power data of S-2, and 96 points of power data of S-3, feature values ​​are constructed to obtain data samples of clear sky day S as clear sky samples.

[0116] In other embodiments, as described above, corresponding power benchmark data can be extracted based on the power data of a clear sky day. Therefore, a clear sky sample can be constructed based on the clear sky meteorological data of a specified clear sky day and the power data of a second historical related day that are relevant to predicting the daily maximum power value of the clear sky day.

[0117] For example, the 96-point meteorological data of the clear sky day S obtained by the above sampling, and the 96-point power data of the three days S-1, S-2, and S-3 are taken as examples. The daily maximum power value of the clear sky day S can be used as the prediction target, and the characteristic value is constructed according to the maximum power value of each of the three days S-1, S-2, and S-3 and the average power value of each daytime period (period when the power generation value is greater than zero), as well as the maximum irradiance, maximum temperature, minimum temperature, average temperature, average humidity, weather type and other information of the measured clear sky day S to obtain the corresponding clear sky sample.

[0118] It should be noted that the method of constructing characteristic values ​​based on the clear sky type, clear sky meteorological data and power data of the second historical related day can be determined according to actual application requirements, etc., and is not specifically limited in this specification.

[0119] In some embodiments, the clear sky meteorological data includes second irradiance data and second temperature data corresponding to the maximum value in the clear sky power data of the third clear sky day. Constructing feature values ​​based on the clear sky type, the clear sky meteorological data, and the power data of the second historical related day may include:

[0120] Acquire preset operating temperature data of the photovoltaic power generation component and second operating temperature data corresponding to the maximum value in the clear sky power data.

[0121] The second blended irradiance data is determined based on the second irradiance data, the second air temperature data, the preset working temperature data and the second working temperature data.

[0122] Characteristic values ​​are constructed according to the clear sky type, clear sky meteorological data, power data of the second historical related day and the second harmonic irradiance data to obtain clear sky samples.

[0123] Among them, the maximum value in the clear sky power data of the third clear sky day is the maximum daily power value of the third clear sky day.

[0124] Specifically, for a specified clear sky day, the maximum value in the clear sky power data can be determined based on the clear sky power data of the clear sky day. According to the collection time corresponding to the maximum value, the irradiation data and temperature data corresponding to the collection time can be determined in the collected clear sky meteorological data of the clear sky day, that is, the second irradiation data and the second temperature data corresponding to the maximum value. Similarly, the operating temperature of the photovoltaic power generation component at the collection time can be obtained, that is, the second operating temperature data corresponding to the maximum value. Based on the second irradiation data, the second temperature data, the preset operating temperature of the photovoltaic power generation component and the second operating temperature data, calculation is performed to obtain the second harmonized irradiation data corresponding to the clear sky day.

[0125] By combining the weather type corresponding to the clear sky day (i.e., clear sky type), meteorological data, power data of the second historical related day, and the second harmonic irradiation data to construct eigenvalues, the eigenvector corresponding to the clear sky day can be obtained as the clear sky sample of the clear sky day.

[0126] Exemplarily, historical meteorological data of historical dates collected by the photovoltaic station are obtained, and the daily measured meteorological data are collected at a frequency of 15 minutes for instantaneous values. For a specified clear sky day S, the past three days S-1, S-2, and S-3 of the clear sky day are used as the second historical related days. According to the collection time corresponding to the maximum value in the power data of the clear sky day S, the corresponding irradiance data and temperature data can be determined in the collected meteorological data of the clear sky day S, and the working temperature of the photovoltaic power generation component at the corresponding time can be collected. Substituting the irradiance data, temperature data, working temperature data, and the preset working temperature of the photovoltaic power generation component into the formula of the aforementioned photovoltaic conversion efficiency physical model, the harmonic irradiance characteristics corresponding to the clear sky day S can be calculated, which is the second harmonic irradiance data.

[0127] Then, the daily maximum power value of the clear sky day S can be used as the prediction target. According to the maximum power value of each of the three days S-1, S-2, and S-3 and the average power value during the daytime period of each day, as well as the measured maximum irradiance, maximum temperature, minimum temperature, average temperature, average humidity, weather type and other information of the clear sky day S, and the characteristics such as the harmonic irradiance of the photovoltaic conversion efficiency physical model as the characteristics of the training data, the characteristic values ​​are constructed to obtain the corresponding clear sky samples.

[0128] It should be noted that, in some cases, the second irradiance data can be considered as the maximum irradiance on the third clear sky day, and the second temperature data can be considered as the maximum temperature on the third clear sky day.

[0129] In some implementations, the weather type corresponding to a non-clear sky day includes the day weather type of a non-clear sky day. Figure 4 As shown, the data sample construction method may further include the following steps.

[0130] S410: Obtain power data of a third historical related day.

[0131] Among them, the third historical relevant day is earlier than the non-clear sky day.

[0132] S420, constructing feature values ​​based on the daily weather type of the non-clear sky day, the non-clear sky meteorological data and the power data of the third historical related day to obtain a second non-clear sky sample for training the third initial prediction model.

[0133] Among them, the daily weather type of a non-clear sky day is a non-clear sky type, and the non-clear sky type refers to a weather type in which factors affecting solar radiation exist, such as cloud cover and fog, and may include at least one of a cloudy weather type, a rainy weather type, a foggy weather type, and the like.

[0134] The third historically relevant date may be the same as or different from the first historically relevant date in this specification.

[0135] The third initial prediction model can be used to predict the photovoltaic power generation on a non-clear sky day. It should be noted that the third initial prediction model is not the same model as the first initial prediction model in this specification.

[0136] Specifically, for the aforementioned cloudy day, the third historical related day can be determined, and the measured power data of the third historical related day can be obtained. A feature vector is constructed according to the daily weather type of the cloudy day (i.e., the corresponding cloudy type), the cloudy meteorological data, and the power data of the third historical related day, and a second cloudy sample of the cloudy day is obtained as a data sample for training a third initial prediction model for predicting photovoltaic power generation on cloudy days.

[0137] It should be noted that for the description of constructing the second non-clear sky sample in this embodiment, please refer to the description of constructing the clear sky sample in this specification, and the details will not be repeated here.

[0138] The method of constructing characteristic values ​​based on the daily weather type of non-clear sky days, non-clear sky meteorological data and power data of the third historical related day can be determined according to actual application requirements, etc., and is not specifically limited in this specification.

[0139] Similar to the first historical related day, the second historical related day and the third historical related day in this specification may be related days that have a greater correlation and reference role in predicting the power of the corresponding day determined by means of data analysis, etc., and are not specifically limited in this specification. In addition, this specification does not specifically limit the structures of the first initial prediction model, the second initial prediction model, and the third initial prediction model.

[0140] In some embodiments, a cloudy day corresponds to cloudy power data; the cloudy meteorological data includes third irradiance data and third temperature data corresponding to the maximum value in the cloudy power data. Constructing a characteristic value based on the daily weather type of the cloudy day, the cloudy meteorological data, and the power data of the third historical related day may include:

[0141] Obtain preset operating temperature data of the photovoltaic power generation component and third operating temperature data corresponding to the maximum value in the non-clear sky power data.

[0142] The third blended irradiance data is determined based on the third irradiance data, the third air temperature data, the preset working temperature data and the third working temperature data.

[0143] Characteristic values ​​are constructed according to the non-clear sky type, non-clear sky meteorological data, power data of the third historical related day and the third harmonic irradiance data to obtain the second non-clear sky sample.

[0144] Among them, the maximum value in the non-clear sky power data is the maximum daily power value on a non-clear sky day.

[0145] It should be noted that, for the description of constructing the second non-clear sky sample in this embodiment, please refer to the description in this specification about constructing characteristic values ​​based on the clear sky type, clear sky meteorological data, power data of the second historical related day and the second harmonic irradiation data to obtain the clear sky sample. The details will not be repeated here.

[0146] This specification provides a method for training a power prediction model. Figure 5 As shown, the power prediction model training method may include the following steps.

[0147] S510, obtaining non-clear sky power data of a non-clear sky day, and a first non-clear sky sample constructed by any of the aforementioned data sample construction methods.

[0148] S520: Input the first non-clear sky sample into the first initial prediction model for prediction to obtain first power prediction data.

[0149] S530. Train a first initial prediction model according to the first power prediction data and the non-clear sky power data to obtain a trained first power prediction model.

[0150] The first power prediction data is the predicted power data of the corresponding non-clear sky day output by the first initial prediction model based on the characteristic data in the input first non-clear sky sample.

[0151] Specifically, for a specified non-clear sky day, the power data of the non-clear sky day, i.e., non-clear sky power data, can be collected. Through the aforementioned data sample construction method, the first non-clear sky sample corresponding to the non-clear sky day can be constructed, wherein the first non-clear sky sample contains the power reference data corresponding to the non-clear sky day, so that the prediction model can learn the influence of factors such as cloud cover on power generation power under non-clear sky weather types based on the solar irradiance variation characteristics under standard clear sky (or, the basic variation characteristics of solar irradiance under the astronomical position of the earth and the sun corresponding to the non-clear sky day). The first non-clear sky sample is used as training data and input into the first initial prediction model that has not completed training to obtain the first power prediction data output by the first initial prediction model.

[0152] The known non-clear sky power data of the non-clear sky day is used as the prediction target, and the parameters of the first initial prediction model are trained and adjusted according to the first power prediction data and the non-clear sky power data, and the loss value of the loss function and other model evaluation criteria, until the accuracy and loss value of the prediction model obtained by the final training meet the conditions, and the trained first power prediction model can be obtained.

[0153] It is understandable that during the model training process, the non-clear sky power data on the non-clear sky day is used as the label data annotated by the first non-clear sky sample.

[0154] In some embodiments, the first initial prediction model may adopt a gradient boosting tree LightGBM model.

[0155] It should be noted that the first initial prediction model may also adopt other model structures according to actual application requirements, etc., and is not specifically limited in this specification.

[0156] In some embodiments, reference Figure 6 As shown, the power prediction model training method may also include the following steps.

[0157] S610, obtaining clear sky power data on a clear sky day, and a clear sky sample constructed by the aforementioned data sample construction method.

[0158] S620. Input the clear sky sample into the second initial prediction model for prediction to obtain second power prediction data.

[0159] S630. Train the second initial prediction model according to the second power prediction data and the clear sky power data to obtain a trained second power prediction model.

[0160] The second power prediction data is the predicted power data of the corresponding clear sky day output by the second initial prediction model based on the characteristic data in the input clear sky sample.

[0161] Specifically, for a specified clear sky day, the power data of the clear sky day, i.e., clear sky power data, can be collected. Through the aforementioned data sample construction method, a clear sky sample corresponding to the clear sky day can be constructed. The clear sky sample is used as training data and input into the second initial prediction model that has not completed training to obtain the second power prediction data output by the second initial prediction model.

[0162] The known clear sky power data of the clear sky day is used as the prediction target, and the parameters of the second initial prediction model are trained and adjusted according to the second power prediction data and the clear sky power data, until the accuracy, loss value, etc. of the prediction model finally trained meet the training termination conditions, and the trained second power prediction model can be obtained.

[0163] During the model training process, the clear sky power data on clear sky days is used as the label data for clear sky samples.

[0164] In some embodiments, the clear sky power data may include the acquired full-day power data of the corresponding clear sky day. During model training, the full-day power data of the clear sky day may be used as a prediction target, so that the trained second power prediction model can predict the full-day power data of the to-be-predicted day with a clear sky weather type.

[0165] In other embodiments, the clear sky power data includes the maximum power data of the corresponding clear sky day. During model training, the maximum power data of the clear sky day can be used as a prediction target, so that the trained second power prediction model can predict the maximum power of the to-be-predicted day with a clear sky weather type.

[0166] In some embodiments, training the second initial prediction model based on the second power prediction data and the clear sky power data may include: training the second initial prediction model based on the maximum value of the second power prediction data and the clear sky power data to obtain the second power prediction model.

[0167] In this embodiment, the second initial prediction model is used to predict the maximum power of a clear sky day corresponding to the clear sky sample. The second power prediction model may also be called a daily power maximum value prediction model or a daily maximum power prediction model.

[0168] In some cases, as described above, for a clear sky day, by normalizing the power data of the clear sky day, the clear sky standard curve of the clear sky day can be obtained, that is, the power benchmark data of the clear sky day. Since the power curve under clear sky conditions is mainly determined by the astronomical law of the relative position of the earth and the sun, in the absence or ignoring of the influence of clouds, the shape of the power curve of a clear sky day can be described by the clear sky standard curve extracted above, and the specific numerical values ​​of each point on the curve can be controlled by the daily maximum power prediction result. Therefore, clear sky samples can be used to train a prediction model for predicting the daily maximum power.

[0169] Specifically, for a specified clear sky day, the maximum value in the known clear sky power data of the clear sky day is used as the prediction target, so as to train and adjust the parameters of the second initial prediction model according to the second power prediction data and the clear sky power data, until the accuracy, loss value, etc. of the prediction model finally trained meet the training termination conditions, and the trained second power prediction model can be obtained.

[0170] It should be noted that the second initial prediction model may adopt a regression model or other model structures according to actual application requirements, etc., and is not specifically limited in this specification.

[0171] In some embodiments, the second initial prediction model includes a first sub-model and a second sub-model. Inputting the clear sky sample into the second initial prediction model for prediction to obtain the second power prediction data may include:

[0172] The clear sky samples are input into the first sub-model for prediction to obtain the power prediction data output by the first sub-model.

[0173] The clear sky sample is input into the second sub-model for prediction to obtain power prediction data output by the second sub-model; wherein the second power prediction data includes the power prediction data output by the first sub-model and the power prediction data output by the second sub-model.

[0174] Accordingly, training the second initial prediction model according to the second power prediction data and the clear sky power data to obtain the trained second power prediction model may include:

[0175] The first sub-model is trained according to the power prediction data and clear sky power data output by the first sub-model to obtain a first trained sub-model.

[0176] The second sub-model is trained according to the power prediction data and clear sky power data output by the second sub-model to obtain a second trained sub-model; wherein the second power prediction model includes the first trained sub-model and the second trained sub-model.

[0177] Among them, the first sub-model and the second sub-model are each used to output corresponding prediction data separately according to the input data sample.

[0178] Specifically, for a specified clear sky day, the clear sky sample of the clear sky day is used as training data and input into the first sub-model and the second sub-model that have not completed training, respectively, to obtain the power prediction data output by the first sub-model for the clear sky sample, and the power prediction data output by the second sub-model for the clear sky sample.

[0179] The clear sky power data of the clear sky day is used as the prediction target, and the parameters of the first sub-model are trained and adjusted according to the power prediction data and the clear sky power data output by the first sub-model, and the parameters of the second sub-model are trained and adjusted according to the power prediction data and the clear sky power data output by the second sub-model, until the training termination conditions are met respectively, to obtain the trained first sub-model and the second sub-model.

[0180] Furthermore, when performing power prediction inference using a second power prediction model including a first trained sub-model and a second trained sub-model, the prediction results of the first trained sub-model and the second trained sub-model can be weighted to obtain a weighted prediction result as the prediction result finally output by the model.

[0181] In some embodiments, the first sub-model may be a ridge regression model, and the second sub-model may be a K-Nearest Neighbours (KNN) model.

[0182] It can be understood that the first sub-model and the second sub-model can each be used to predict the daily maximum power value, that is, the first sub-model and the second sub-model can both be daily maximum power prediction sub-models.

[0183] It should be noted that the structures of the first sub-model and the second sub-model can be determined according to actual application requirements, etc., and are not specifically limited in this specification.

[0184] In some embodiments, reference Figure 7a As shown, the power prediction model training method may also include the following steps.

[0185] S710. Obtain a second non-clear sky sample constructed by the aforementioned data sample construction method.

[0186] S720. Input the second non-clear sky sample into the third initial prediction model for prediction to obtain third power prediction data.

[0187] S730. Train a third initial prediction model according to the third power prediction data and the non-clear sky power data to obtain a trained third power prediction model.

[0188] The third power prediction data is the predicted power data of the corresponding non-clear sky day output by the third initial prediction model based on the characteristic data in the input second non-clear sky sample.

[0189] Specifically, for a specified cloudy day, the second cloudy sample corresponding to the cloudy day can be constructed by the aforementioned data sample construction method. The second cloudy sample is used as training data and input into the third initial prediction model that has not completed training to obtain the third power prediction data output by the third initial prediction model.

[0190] The known non-clear sky power data of the non-clear sky day is used as the prediction target, and the parameters of the third initial prediction model are trained and adjusted according to the third power prediction data and the non-clear sky power data until the training termination condition is met, so as to obtain the trained third power prediction model.

[0191] In some embodiments, the non-clear sky power data may include the acquired whole-day power data of the corresponding non-clear sky day. Accordingly, during model training, the whole-day power data of the non-clear sky day may be used as a prediction target.

[0192] In other embodiments, the non-clear sky power data includes the maximum power data of the corresponding non-clear sky day. Accordingly, during model training, the maximum power data of the non-clear sky day can be used as a prediction target.

[0193] Furthermore, in the stage of predicting and reasoning the power data of the predicted day, if the predicted day is not clear sky weather, the power prediction result output by the first power prediction model can be corrected according to the output result of the third power prediction model.

[0194] It should be noted that the structures of the first initial prediction model, the second initial prediction model and the third initial prediction model can be determined according to actual application requirements, etc., and are not specifically limited in this specification.

[0195] In some embodiments, the third power prediction model and the second power prediction model are the same power prediction model. The second power prediction model is obtained by training the second initial prediction model based on the second power prediction data and clear sky power data of a clear sky day; the second power prediction data is obtained by predicting the second initial prediction model based on the input clear sky sample of a clear sky day.

[0196] In this embodiment, the third initial prediction model and the second initial prediction model are the same initial prediction model.

[0197] Specifically, the clear sky sample and the second non-clear sky sample constructed by the aforementioned data sample construction method can be used to train the same second initial prediction model (or, the third initial prediction model), so that the trained second power prediction model (or, the third power prediction model) has the function of predicting the power data of the predicted day of any weather type.

[0198] It can be understood that, when the third power prediction model and the second power prediction model are the same power prediction model, the clear sky sample is constructed in the same manner as the second non-clear sky sample.

[0199] In some embodiments, the third power prediction model and the second power prediction model are the same daily maximum power prediction model.

[0200] Exemplarily, historical meteorological data of historical dates collected by the photovoltaic station are obtained, and the daily measured meteorological data are collected at a frequency of 15 minutes. For a clear sky day S, the corresponding irradiance data and temperature data are determined according to the collection time corresponding to the maximum value in the power data of the clear sky day S, and the irradiance data, temperature data and operating temperature data, as well as the preset operating temperature of the photovoltaic power generation component are substituted into the formula of the aforementioned photovoltaic conversion efficiency physical model, and the second harmonic irradiance data corresponding to the clear sky day S can be calculated.

[0201] Then, the daily maximum power value of the clear sky day S can be used as the prediction target. According to the maximum power value of each of the three days S-1, S-2, and S-3 and the average power value during the daytime period of each day, as well as the measured maximum irradiance, maximum temperature, minimum temperature, average temperature, average humidity, weather type and other information of the clear sky day S, and the characteristics such as the harmonic irradiance of the photovoltaic conversion efficiency physical model as the characteristics of the training data, the characteristic values ​​are constructed to obtain the corresponding clear sky samples.

[0202] Similarly, for cloudy day D, the corresponding irradiance data and temperature data are determined according to the collection time corresponding to the maximum value in the power data of cloudy day D, and the irradiance data, temperature data and operating temperature data, as well as the preset operating temperature of the photovoltaic power generation component, are calculated to obtain the third harmonized irradiance data corresponding to cloudy day D.

[0203] Then, the daily maximum power value of the cloudy day D can be used as the prediction target. The feature value is constructed based on the maximum power value of each of the three days D-1, D-2, and D-3 and the average power value during the daytime period of each day, as well as the measured maximum irradiance, maximum temperature, minimum temperature, average temperature, average humidity, weather type and other information of the cloudy day D, and the harmonic irradiance and other features as the features of the training data to obtain the corresponding second cloudy sample.

[0204] The clear sky samples and their corresponding daily maximum power values, as well as the second non-clear sky samples and their corresponding daily maximum power values, are used to train the same initial prediction model to obtain a trained daily maximum power prediction model.

[0205] Further, the second power prediction model may include a first trained sub-model and a second trained sub-model. In some embodiments, the first trained sub-model adopts a ridge regression model, and the second trained sub-model adopts a KNN model.

[0206] For example, reference Figure 7b As shown, the process of the daily maximum power prediction algorithm may include:

[0207] In the model training phase, the photovoltaic power generation data and meteorological data collected by the photovoltaic station on each historical date are obtained, and the power data and meteorological data are preprocessed, such as outlier processing and missing value processing. Among them, the timestamp of the measured meteorological data corresponds to the timestamp of the power data. The preprocessed power data and measured meteorological data are aggregated on a daily basis, and the daily maximum power is aggregated as the prediction target of the daily maximum power prediction model.

[0208] Features are constructed for a certain date, including conventional meteorological features, power hysteresis features, and harmonic irradiance features considering conversion efficiency. Conventional meteorological features include the maximum irradiance, maximum temperature, minimum temperature, average temperature, average humidity, weather type, and other information of the measured meteorological data on that date; power hysteresis features include the maximum power value of each day and the average power value during the daytime (power generation value is greater than zero) in the three days before that date; and the harmonic irradiance features considering conversion efficiency are the harmonic irradiance features combined with the physical model of photovoltaic conversion efficiency. In this way, data samples for each date are obtained.

[0209] According to the constructed data samples of each historical date, the daily maximum power value corresponding to each date is taken as the prediction target, and the daily maximum power prediction model is trained, that is, the ridge regression model and the KNN model are trained respectively to obtain the trained ridge regression model and KNN model.

[0210] In the inference stage, the weather forecast data of the day to be predicted is obtained from an external meteorological source and data preprocessing is performed. According to the above method, feature construction is performed based on the preprocessed weather forecast data and the daily power data of the three days before the day to be predicted, and the feature data of the day to be predicted is obtained, which is input into the trained daily power maximum prediction model for inference prediction, that is, the constructed feature data is respectively input into the ridge regression model and the KNN model for inference prediction. A weight parameter is used to perform a weighted average calculation on the prediction results output by the ridge regression model and the KNN model to obtain the final predicted daily maximum power value of the day to be predicted. Among them, the weight parameter can be determined according to actual application requirements.

[0211] This specification provides a power prediction method. Figure 8 As shown, the power prediction method may include the following steps.

[0212] S810: Obtain the weather forecast data for the day to be predicted and the second power benchmark data for the same period in history.

[0213] Among them, the meteorological forecast data includes the forecast weather type corresponding to the day to be predicted; the forecast weather type includes the forecast weather type for the whole day; the date of the same period in history and the date of the day to be predicted differ by an integer number of years; the second power benchmark data is used to describe the regularity of photovoltaic power generation under the astronomical position of the earth and the sun corresponding to the day to be predicted and under weather conditions without cloud cover.

[0214] S820: If the weather type forecast for the entire day is a non-clear sky type, obtain power data for the first historical day; wherein the first historical day is earlier than the day to be predicted.

[0215] S830: Construct first weather characteristic data based on the forecast weather type, the meteorological forecast data, the power data of the first historical day and the second power reference data.

[0216] S840: Input the first weather characteristic data into a first power prediction model trained by any of the aforementioned power prediction model training methods to obtain a first power prediction result for the day to be predicted.

[0217] The day to be predicted is a day for which the photovoltaic power generation power needs to be predicted.

[0218] The historical contemporaneous date is a day in the historical year that is the same as the date of the day to be predicted. For example, if the day to be predicted is October 23, 2024, the historical contemporaneous date can be October 23, 2023 or October 23, 2022, etc.

[0219] The forecast weather type may include the overall weather type of the forecasted day to be predicted, and may also include the weather types of different time periods of the forecasted day to be predicted.

[0220] Meteorological forecast data may include commonly used daily weather forecast information, such as daily maximum temperature, minimum temperature, average wind speed, average humidity, weather type (sunny, cloudy, etc.), etc., as well as numerical weather forecast information, such as 15-minute solar irradiance, temperature, humidity, wind speed, wind direction, air pressure, weather type and other field information.

[0221] The first weather characteristic data is a feature vector to be predicted of the day to be predicted obtained by constructing features based on the predicted weather type, the meteorological forecast data, the power data of the first historical day, and the second power reference data.

[0222] The first power prediction result is the predicted power curve of the day to be predicted.

[0223] In some cases, the same date in different years corresponds to the same or similar astronomical position of the earth and the sun, and the regularity of photovoltaic power generation on two days of the same date under the weather conditions without cloud cover is the same or similar, that is, the same date in different years corresponds to the same power benchmark data. Therefore, according to the date of the day to be predicted, the historical same period day can be determined, and the power benchmark data of the historical same period day can be used as the power benchmark data of the day to be predicted.

[0224] Specifically, for the day to be predicted for which power needs to be predicted, the weather forecast data for the day to be predicted can be obtained from the relevant weather source. According to the date of the day to be predicted, the historical date of the same date can be determined from the historical years, and the second power reference data corresponding to the historical date can be obtained as the power reference data for the day to be predicted.

[0225] If it is determined according to the meteorological forecast data that the forecast weather type for the day to be predicted is a non-clear sky type, such as any of the weather types such as cloudy, rainy, sleeting, foggy, etc., then the power data of the first historical day is obtained according to the first historical day earlier than the day to be predicted. Feature values ​​are constructed based on the forecast weather type, meteorological forecast data, power data of the first historical day, and second power reference data to construct the first weather characteristic data of the day to be predicted. The first weather characteristic data is input into the first power prediction model trained by the aforementioned power prediction model training method to obtain the predicted power of the day to be predicted output by the model according to the first weather characteristic data, which is the first power prediction result.

[0226] It should be noted that for the description of constructing characteristic values ​​based on the forecast weather type, meteorological forecast data, power data of the first historical day and the second power benchmark data, please refer to the description of constructing the first non-clear sky sample in this specification, and for the description of the selection method of the first historical day, please refer to the description of the first historical related day in this specification. The details will not be repeated here.

[0227] In some embodiments, according to the above description of the method for obtaining the first power reference data and the description of the method for obtaining the clear sky standard curve for clear sky days, it can be understood that based on the subset of clear sky samples extracted from the historical power data for more than one year (i.e., the power data set for clear sky days), the clear sky standard curve corresponding to each date in a year (i.e., the power reference data corresponding to each date) can be summarized and stored according to the date. When predicting the power of the predicted day, the clear sky standard curve corresponding to the date can be found according to the date of the predicted day, which is the second power reference data for the same period in history.

[0228] In other embodiments, after determining the historical same period day, the method for obtaining the second power reference data can be determined based on the known weather type of the historical same period day. If the weather type of the historical same period day is a clear sky type, the second power reference data can be obtained by normalizing the power data of the historical same period day; if the weather type of the historical same period day is a non-clear sky type, the second power reference data can be obtained based on the power data of the clear sky day adjacent to the historical same period day. For a specific description, please refer to the description of the method for obtaining the first power reference data in this specification, which will not be repeated here.

[0229] In some implementations, the power prediction method may further include:

[0230] If the weather type for the entire day is predicted to be clear sky type, the power data of the second historical day is obtained; wherein the second historical day is earlier than the day to be predicted.

[0231] The second meteorological characteristic data are constructed based on the clear sky type, meteorological forecast data and power data of the second historical day.

[0232] The second weather characteristic data is input into the second power prediction model trained by the aforementioned power prediction model training method to obtain the second power prediction result for the day to be predicted.

[0233] The second weather characteristic data is a feature vector obtained by constructing features based on the clear sky type, weather forecast data, and power data of the second historical day.

[0234] The second power prediction result may be a predicted power curve of the day to be predicted.

[0235] Specifically, if the forecast weather type for the day to be predicted is determined to be a clear sky type according to the meteorological forecast data, the power data of the second historical day is obtained according to the second historical day earlier than the day to be predicted. Feature values ​​are constructed based on the clear sky type, the meteorological forecast data, and the power data of the second historical day to construct the second weather characteristic data of the day to be predicted. The second weather characteristic data is input into the second power prediction model trained by the aforementioned power prediction model training method to obtain the predicted power output by the model, which is the second power prediction result.

[0236] It should be noted that for the description of constructing the second meteorological characteristic data based on the clear sky type, weather forecast data and the power data of the second historical day, please refer to the description of clear sky sample construction in this manual, and for the description of the selection method of the second historical day, please refer to the description of the second historical related day in this manual. The details will not be repeated here.

[0237] In some embodiments, reference Figure 9a As shown, inputting the second weather characteristic data into the second power prediction model trained by the aforementioned power prediction model training method to obtain the second power prediction result of the day to be predicted may include the following steps.

[0238] S910. Input the second weather characteristic data into the second power prediction model to obtain the clear sky power prediction result output by the power prediction model; wherein the clear sky power prediction result is the maximum power data of the predicted day.

[0239] S920. Perform numerical calculation based on the clear sky power prediction result and the second power reference data to obtain a second power prediction result.

[0240] In this embodiment, the second power prediction model is used to predict the maximum power of the day to be predicted.

[0241] Specifically, the second weather characteristic data is input into the second power prediction model, so that the second power prediction model predicts the maximum power of the day to be predicted, and the maximum power prediction data of the day output by the model is obtained, which is the clear sky power prediction result. According to the clear sky power prediction result and the acquired second power reference data, numerical calculation is performed to obtain the second power prediction result, which is the final power prediction result when the day to be predicted is clear sky weather.

[0242] In some embodiments, the ratio between the clear sky power prediction result and the maximum value in the second power reference data can be calculated, and then the ratio and other values ​​in the second power reference data are used to perform calculations to obtain the second power prediction result.

[0243] In other embodiments, the second power reference data is a normalized power curve (ie, a normalized clear sky standard curve).

[0244] Exemplarily, the second power reference data includes 96 points of normalized power data. Figure 9b As shown, the process of the power prediction algorithm under the clear sky classification may include: performing numerical restoration according to the clear sky standard curve of the day to be predicted and the daily maximum power prediction result (clear sky power prediction result) to obtain a 96-point power prediction result of the day to be predicted.

[0245] Further, exemplarily, the second power benchmark data includes 96 points of normalized power data. The second power prediction model may include a trained ridge regression model and a KNN model. The ridge regression model can better learn the linear relationship between the input features and the output targets, and the KNN model can search for target variable predictions of similar samples from historical data. The combination of the two can maximize the prediction accuracy. The second meteorological characteristic data is input into the ridge regression model and the KNN model respectively to obtain the daily maximum power prediction results output by the ridge regression model and the KNN model respectively, and then a weight parameter can be used to perform a weighted average on the results of the two to obtain a clear sky power prediction result with higher accuracy, thereby controlling the midday peak size of the 96-point prediction curve (it is generally believed that the daily maximum power value is in the midday period).

[0246] In the above implementation, when the day to be predicted is clear sky weather, the power curve of the day to be predicted under clear sky conditions is predicted by utilizing the clear sky standard curve characteristics corresponding to the day to be predicted extracted based on historical data and the daily maximum power value prediction model, which can effectively improve the prediction accuracy of photovoltaic power generation under clear sky conditions.

[0247] In some embodiments, if the forecast weather type for the whole day is a non-clear sky type, refer to Fig.10 As shown, the power prediction method may further include the following steps.

[0248] S1010, obtaining power data of the third historical day, and determining a preset weather type of the day to be predicted; wherein the third historical day is earlier than the day to be predicted.

[0249] S1020: Construct third weather characteristic data based on the preset weather type, weather forecast data, and power data of the third historical day.

[0250] S1030. Input the third weather characteristic data into the second power prediction model trained by the aforementioned power prediction model training method to obtain a power prediction reference result corresponding to a preset weather type.

[0251] S1040: Correct the first power prediction result based on the power prediction reference result to obtain a corrected power prediction result for the day to be predicted.

[0252] Among them, the preset weather type is the weather type set for the day to be predicted, which can be any of the weather types such as sunny, cloudy, rainy, rainy and snowy, foggy and haze, and can be the same as the weather type forecast for the whole day.

[0253] The power forecast reference result is a forecast power curve corresponding to a preset weather type, and the revised power forecast result is a revised forecast power curve for the day to be forecasted.

[0254] It can be understood that, in this embodiment, the third weather characteristic data is constructed in the same manner as the aforementioned second weather characteristic data.

[0255] Specifically, if the forecast weather type for the day to be predicted is a non-clear sky type, the power data of the third historical day can also be obtained. The weather type field of the day to be predicted is set, the preset weather type is determined, and the characteristic value is constructed based on the preset weather type, the meteorological forecast data, and the power data of the third historical day to obtain the third weather characteristic data of the day to be predicted.

[0256] The third weather characteristic data is input into the second power prediction model trained by the aforementioned power prediction model training method (the second power prediction model and the aforementioned third power prediction model are the same model), and the power prediction result corresponding to the preset weather type output by the second power prediction model is obtained, which is the power prediction reference result. The power prediction reference result is used to correct the first power prediction result obtained to obtain a corrected power prediction result, which is the final power prediction result when the predicted day is non-clear sky weather.

[0257] It should be noted that for the description of constructing the third weather characteristic data by using the preset weather type, meteorological forecast data and the power data of the third historical day, please refer to the description of constructing the second non-clear sky sample in this manual, and for the description of the selection method of the third historical day, please refer to the description of the third historical related day in this manual. The details will not be repeated here.

[0258] In some embodiments, the preset weather type is the forecast all-day weather type. Based on the power prediction reference result corresponding to the forecast all-day weather type, the degree of deviation between the first power prediction result and the power prediction reference result can be calculated, so that the first power prediction result can be corrected according to the degree of deviation.

[0259] Considering that for cloudy, rainy and other non-clear sky weather types, although the clear sky standard curve feature (i.e., power benchmark data) is added to the first non-clear sky sample, the corresponding model may weaken the influence of this feature during the training and learning stage, resulting in the first power prediction model after training to predict the power of non-clear sky days. When inferring, it will obtain some prediction points that deviate from the regularity presented by the corresponding clear sky standard curve. Therefore, it is necessary to perform forced post-processing based on this feature again to adjust the prediction points that do not conform to the regularity of the curve.

[0260] For example, if the forecast weather type for the entire day of the predicted day P is cloudy, the preset weather type can also be cloudy. The day P-1 before the predicted day P is taken as the third historical day, and the characteristic value is constructed based on the cloudy type, the weather forecast data and the power data of the day P-1 to construct the weather characteristic data of the predicted day P.

[0261] The constructed weather characteristic data is input into the daily maximum power prediction model including the ridge regression model and the KNN model, and the daily maximum power value after weighted average of the model output can be obtained. According to the daily maximum power value and the power benchmark data of the predicted day P, the predicted power curve corresponding to the cloudy weather type is obtained, which is the power prediction reference result corresponding to the preset weather type.

[0262] According to the power prediction reference result and the first power prediction result of the predicted day P obtained by the power prediction method described above, the degree of deviation between the two can be calculated. If the degree of deviation exceeds a certain threshold, the first power prediction result can be corrected according to the power prediction reference result. For example, any one of the interpolation method, linear regression, etc. can be used to correct it to obtain a corrected power prediction result. In this way, forced post-processing based on the characteristics of the clear sky standard curve is achieved to adjust the prediction points that do not conform to the regularity of the curve.

[0263] In other embodiments, the number of preset weather types is multiple. According to the weather types corresponding to different time periods of the day to be predicted, the reference power data corresponding to different time periods in the first power prediction result can be determined according to the power prediction reference result, so that the first power prediction result can be corrected according to the reference power data corresponding to different time periods.

[0264] In some embodiments, the forecast weather type includes the weather type predicted for any period of the day to be predicted. Fig.11 As shown, the first power prediction result is corrected based on the power prediction reference result to obtain the corrected power prediction result of the day to be predicted, which may include the following steps.

[0265] S1110. Determine time period power reference data corresponding to the weather type of any time period according to the power prediction reference result.

[0266] S1120 , according to the time sequence of the multiple time periods of the day to be predicted, splicing the time period power reference data corresponding to each of the multiple time periods to obtain spliced ​​power data.

[0267] S1130, performing calculations based on the spliced ​​power data and the preset fault tolerance data to obtain fault tolerance correction data.

[0268] S1140. Correct the first power prediction result according to the fault-tolerant correction data to obtain a corrected power prediction result.

[0269] The weather type of any period can be any of sunny, cloudy, rainy, foggy, etc. It should be noted that if the weather type of a period is cloudy, then the collection time points in the period (corresponding to the collection time points in the aforementioned data sample construction method) all correspond to cloudy weather types.

[0270] In this embodiment, if the number of preset weather types is multiple, then the number of power prediction reference results is multiple. The timestamp of any power prediction reference result is the same as the timestamp of the first power prediction result. That is, assuming that the first power prediction result includes the predicted power values ​​at 96 time points of the day to be predicted, then any power prediction reference result also includes the predicted power values ​​at the corresponding 96 time points.

[0271] The preset fault tolerance data is data preset to calculate the fault tolerance range of the first power prediction result, and may be a percentage value.

[0272] The error-tolerance correction data may be used to limit the error-tolerance range of the first power prediction result. The error-tolerance correction data is also referred to as a prediction interval or a confidence interval in this specification.

[0273] Specifically, according to a plurality of preset weather types, third weather characteristic data corresponding to each preset weather type are respectively constructed and respectively input into the second power prediction model to obtain power prediction reference results corresponding to each preset weather type. According to the weather type of a certain period of time on the day to be predicted, a section of data corresponding to the period is determined in the power prediction reference result corresponding to the corresponding preset weather type, which is the corresponding period power reference data. According to this method, the period power reference data corresponding to a plurality of period of time on the day to be predicted can be determined.

[0274] Then, according to the time sequence of the multiple time periods, the corresponding time period power reference data are spliced ​​to obtain spliced ​​power data. According to the spliced ​​power data and the preset fault tolerance data, calculation is performed to obtain fault tolerance correction data for indicating the fault tolerance range, and the first power prediction result is corrected according to the fault tolerance correction data to obtain a corrected power prediction result, which is the power prediction data of the day to be predicted that is finally obtained.

[0275] For example, if the forecast weather type for the forecast day P is cloudy, the preset weather types can be sunny, cloudy and rainy. The day before the forecast day P-1 is taken as the third historical day, and the characteristic value is constructed based on the sunny type, the weather forecast data and the power data of the day P-1 to construct the weather characteristic data of the forecast day P.

[0276] The constructed weather characteristic data is input into the daily maximum power prediction model including the ridge regression model and the KNN model, and the daily maximum power value after weighted average of the model output can be obtained. According to the daily maximum power value and the power reference data of the predicted day P, the power prediction curve corresponding to the preset sunny weather type is calculated, which is the power prediction reference result corresponding to the sunny weather type. According to this method, the power prediction curve corresponding to the cloudy weather type and the power prediction curve corresponding to the rainy weather type can be obtained respectively.

[0277] Assuming that the weather type in the morning of the predicted day is cloudy, the power curve corresponding to the morning period is obtained according to the power prediction curve corresponding to the cloudy weather type; assuming that the weather type in a certain two hours in the afternoon is sunny, the power curve corresponding to the sunny weather type is obtained according to the power prediction curve corresponding to the sunny weather type. Similarly, the power curves corresponding to multiple time periods can be obtained. According to the time sequence between multiple time periods, the corresponding power curves are spliced ​​to obtain the spliced ​​power curve, which is the spliced ​​power data.

[0278] According to the spliced ​​power curve and the preset fault tolerance data, two curves for limiting the first power prediction result can be calculated as fault tolerance correction data. According to the fault tolerance correction data, the first power prediction result can be corrected by calculating the degree of deviation between the first power prediction result and the fault tolerance correction data, and then using any one of the methods such as interpolation or linear regression according to the relationship between the degree of deviation and the preset deviation threshold, or other correction methods can be used to correct the first power prediction result.

[0279] It should be noted that, in this specification, the power prediction reference result corresponding to the preset weather type is also referred to as the standard prediction curve corresponding to the preset weather type. The spliced ​​power data is also referred to as the standard curve of the day to be predicted. The process of correcting the first power prediction result according to the fault-tolerant correction data can also be referred to as the post-processing process.

[0280] In some implementations, the to-be-predicted day includes multiple to-be-predicted time points; the first power prediction result includes power data at any to-be-predicted time point; and the fault-tolerant correction data includes correction data at any to-be-predicted time point. Correcting the first power prediction result according to the fault-tolerant correction data may include:

[0281] If the power data at any to-be-predicted time point is within the range of the correction data at the corresponding to-be-predicted time point, the power data at the corresponding to-be-predicted time point is kept unchanged.

[0282] If the power data at any to-be-predicted time point is greater than the maximum value of the corrected data at the corresponding to-be-predicted time point, the power data at the corresponding to-be-predicted time point is corrected to the maximum value of the corrected data at the corresponding to-be-predicted time point.

[0283] If the power data at any to-be-predicted time point is less than the corrected data at the corresponding to-be-predicted time point, the power data at the corresponding to-be-predicted time point is corrected to the minimum value of the corrected data at the corresponding to-be-predicted time point.

[0284] The time point to be predicted is the time point at which the corresponding power needs to be predicted.

[0285] It can be understood that the spliced ​​power data includes a power reference value at any time point to be predicted.

[0286] Specifically, for the specified time point to be predicted on the day to be predicted, calculation can be performed based on the power reference value corresponding to the time point in the spliced ​​power data and the preset fault-tolerant data to obtain the corrected data corresponding to the time point, thereby obtaining the fault-tolerant corrected data based on the corrected data corresponding to each of the multiple time points to be predicted.

[0287] When the first power prediction result is corrected using fault-tolerant correction data, for a certain time point to be predicted, if the power data corresponding to the time point in the first power prediction result is within the range of the correction data corresponding to the time point, the power data at the time point remains unchanged.

[0288] If the power data corresponding to the time point in the first power prediction result is greater than the maximum value of the corrected data corresponding to the time point, the power data at the time point in the first power prediction result is corrected to the maximum value of the corresponding corrected data.

[0289] If the power data corresponding to the time point in the first power prediction result is less than the minimum value of the corrected data corresponding to the time point, the power data at the time point in the first power prediction result is corrected to the minimum value of the corresponding corrected data.

[0290] According to the above method, the power data at each time point to be predicted in the first power prediction result can be corrected to obtain a corrected power prediction result.

[0291] Exemplarily, similar to the sampling time points of 96 meteorological data obtained by uniformly sampling at a sampling interval of 15 minutes in the aforementioned data sample construction method, the predicted day includes 96 time points to be predicted, and the first power prediction result is a prediction curve including power values ​​of corresponding 96 points, and the spliced ​​power data is a standard curve including power reference values ​​of corresponding 96 points after splicing.

[0292] The error tolerance percentage δ is introduced into the above standard curve as the preset error tolerance data to calculate the correction data corresponding to each time point to be predicted (hereinafter referred to as the confidence interval of each point). The specific process is: for the point x at time t in the original standard curve t , the confidence interval that can be accepted is calculated as:

[0293] [(1-δ)x t ,(1+δ)x t ]

[0294] For each point in the original 96-point prediction curve (first power prediction result) of non-clear sky weather types such as cloudy and rainy weather output by the prediction model The confidence interval [(1-δ)x t ,(1+δ)x t ] to correct the original prediction result:

[0295]

[0296] Exemplarily, non-clear sky types include cloudy types, rainy and extreme weather types. The photovoltaic power generation short-term power prediction method considering clear sky laws provided in this specification includes a model training phase and a model reasoning phase. Fig.12a As shown, the process of the model training phase may include:

[0297] S01. Measured data collection. Obtain the data required for the power prediction algorithm, including historical photovoltaic power data and meteorological data collected by photovoltaic stations. Among them, the historical power is collected at a frequency of 15 minutes for instantaneous values, and the daily power generation data at 96 points is obtained by uniform sampling at a sampling interval of 15 minutes from 00:15 to 00:00 the next day. The measured meteorological data is collected by environmental monitoring instruments and solar irradiation instruments, including field information such as horizontal total irradiation, ambient temperature, ambient humidity, wind speed, wind direction, air pressure, and weather type. The timestamp of the measured meteorological data corresponds one-to-one to the timestamp of the power data.

[0298] S02, data preprocessing. First, data cleaning and outlier processing are performed. Due to instrument collection and other reasons, measured data is prone to produce outliers, missing values ​​and other abnormal information. Perform data distribution analysis on each field (power, meteorological field) to identify and remove outliers. Then perform missing value processing. After removing outliers, the corresponding outliers will become missing value points. Linear interpolation and other methods can be used to interpolate missing values ​​for these points together with the points originally missing during the collection process. Taking the linear interpolation method as an example, for the value z of the missing value point t t , can be based on the value z of the previous point t-1 t-1 and the value z of the next point t+1 t+1 And the coefficient α is interpolated as follows:

[0299] z t =α·z t-1 +(1-α)·z t+1

[0300] The value of the coefficient α can be determined according to actual application requirements, etc.

[0301] S03, extraction of clear sky standard curve. For the 96-point power curve of each day in the historical power data, the power curve of the date with clear sky type is filtered out according to the historical weather type to obtain a sunny day sample subset, and the corresponding date information is saved. The clear sky power curve in the above sunny day sample subset is normalized, that is, taking the day as the smallest unit, the 96-point power value of each day is divided by the maximum power value of the day to process the daily power curve value to a value between 0 and 1, and obtain the daily clear sky standard curve. Through visual analysis, it is found that these clear sky standard curves have the characteristics of periodic changes, and are consistent with the cognition of the law of change of the relative position of the earth's orbit and the sun (that is, the law of change of the astronomical position of the earth and the sun). The clear sky standard curve of a certain day can reflect the shape characteristics of the photovoltaic power generation power curve under pure clear sky conditions without cloud cover on that day.

[0302] Based on the sunny sample subset extracted from historical data of more than one year, the clear sky standard curve corresponding to each date of the year can be summarized and extracted. Among them, the dates included in the sunny sample subset (i.e., the dates with clear sky weather type) correspond to the clear sky standard curve obtained after normalization according to the power curve of the corresponding date. For a date not included in the sunny sample subset (i.e., the date with non-clear sky weather type), the clear sky day adjacent to the not included date can be determined from the above-mentioned included dates, and the clear sky standard curve of the adjacent clear sky day can be linearly interpolated to obtain the clear sky standard curve of the date. For the specific interpolation process, please refer to the description of the method for obtaining the first power reference data in this manual, which will not be repeated here.

[0303] S04. Construction of features for the daily maximum power prediction model. Aggregate the pre-processed 15-minute frequency power data and measured meteorological data of each date on a daily basis, and aggregate the daily maximum power value as the prediction target of the daily maximum power prediction model. For a certain date, combine the maximum power value of each day and the average power value during the daytime period (the power generation value is greater than zero) of the previous three days, as well as the maximum irradiance, maximum temperature, minimum temperature, average temperature, average humidity, weather type and other information of the measured meteorological information of that date, and combine the harmonic irradiance and other features of the photovoltaic conversion efficiency physical model as the features of the training data to construct data samples. Among them, please refer to the description of the calculation method of the first harmonic irradiance data in this manual for the calculation method of the harmonic irradiance feature, and the details will not be repeated here.

[0304] S05. Training of the daily maximum power prediction model. Based on the daily characteristic variables constructed in the training data of the daily maximum power prediction model described in S04 and the corresponding target daily maximum power value, a ridge regression model and a KNN (K-nearest neighbor) regression model are trained, and a weight parameter is used to weight the average of the two results. At the model level, the ridge regression model can better learn the linear relationship between the input features and the output targets, and the KNN model can search for target variable predictions of similar samples from historical data. The combination of the two can maximize the prediction accuracy and thus control the midday peak size of the 96-point prediction curve.

[0305] S06. Feature construction under cloudy and rainy weather. According to the weather type of the measured meteorological data, determine whether it is sunny or cloudy or rainy. If it is sunny, the training process ends; if it is cloudy or rainy, reconstruct the features for training the 96-point power prediction model (corresponding to the first power prediction model in this specification).

[0306] In cloudy and rainy weather, the influence of cloud-related factors needs to be considered, and the influencing factors are more complicated than those in clear sky conditions. Therefore, the features of this type of weather are reconstructed, and the clear sky standard curve features are added to the training data samples so that the model can learn the morphological features of the power benchmark curve and the influence of clouds respectively.

[0307] refer to Figure 12b As shown, in the training stage of the power prediction algorithm under the cloudy or rainy weather classification, features are constructed based on the preprocessed measured meteorological data and power data, and features are constructed for each point in the 96 points of weather data on cloudy and rainy days as input features of the model. The power data corresponding to each point is used as the prediction target of each point (i.e., as the label of the data at each point) to train the model to predict the power value of each point. The input features of each point include: conventional meteorological features corresponding to the current point, such as irradiance, temperature, humidity, wind speed, weather type, etc.; timestamp features, such as the point number (0-95) and hours corresponding to the current point on the day, and the coded information after the sinusoidal function transformation of the point number and hours, etc.; time series features (power lag features) constructed based on historical power data, such as the power lag value at the same time point corresponding to the current point in the past day (that is, the power value at the same time point at the current point in the past day), and the daily maximum and daytime average power values ​​of the past day, etc.; other features, including the irradiance features after reconciliation using the conversion efficiency model described in S04, and the clear sky standard curve features described in S03, etc. The clear sky standard curve feature corresponding to each point is the value of the above-extracted normalized clear sky standard curve at the corresponding point.

[0308] S07, 96-point prediction model training. The features reconstructed according to the method described in S06 are used as input samples to train the 96-point prediction model.

[0309] Continue to refer Figure 12b As shown, after constructing the features of the historical data under cloudy and rainy weather according to the method described in S6, the gradient boosting tree LightGBM model is used for model training.

[0310] Accordingly, refer to Fig.12c As shown, the process of the model inference prediction stage can include:

[0311] S11, meteorological forecast data collection and preprocessing. The meteorological forecast data for the day to be predicted is obtained from an external meteorological source, including common daily weather forecast information, i.e., daily maximum temperature, minimum temperature, average wind speed, average humidity, weather type (sunny, cloudy, etc.), and numerical weather forecast, i.e., field information such as solar irradiance, temperature, humidity, wind speed, wind direction, air pressure, and weather type every 15 minutes. The timestamp (excluding date information) corresponds to the timestamp of the measured power data described in S01 above, and the field information corresponds to the measured meteorological data field described in S01 above. The meteorological forecast data for the day to be predicted is processed for abnormal values ​​and missing values ​​according to the data preprocessing method described in S02 above.

[0312] S12, obtaining clear sky standard curve: According to the various dates and their corresponding clear sky standard curves stored in S03, obtain the clear sky standard curve corresponding to the same date as the date to be predicted, for use in the subsequent power prediction process.

[0313] S13, daily maximum power prediction model reasoning. In the stage of predicting and reasoning the daily maximum power value of the day to be predicted, according to the feature construction method in S04 above, the feature construction for daily maximum power prediction is performed based on the meteorological forecast data collected in S11, and then the trained daily maximum power prediction model is used for reasoning to obtain the prediction result of the daily maximum power value of the day to be predicted.

[0314] S14, power prediction under the sunny classification. Judge according to the weather type of the day to be predicted. If the weather type of the day to be predicted is a clear sky type, since the power curve under clear sky conditions is mainly determined by the astronomical law of the relative position of the earth and the sun, therefore, in the absence or without considering the influence of clouds, the shape of the power curve can be described by the corresponding clear sky standard curve, and the specific numerical value is controlled by the daily maximum power prediction result. The clear sky standard curve of the day to be predicted obtained in S12 has been normalized to a value between 0 and 1. Therefore, the short-term power prediction model based on the clear sky law obtains a 96-point power curve through the sunny power prediction method. That is, according to the numerical values ​​of the 96 points in the clear sky standard curve, they are multiplied by the daily maximum power prediction value obtained by reasoning in S13 to perform numerical restoration to obtain the 96-point power prediction curve of the day to be predicted and output it.

[0315] S15, power prediction under cloudy and rainy weather classification. If the weather type of the day to be predicted is judged to be cloudy or rainy, the characteristics of cloudy and rainy weather are constructed and input into the short-term power prediction model based on the law of clear sky. The short-term power prediction model based on the law of clear sky uses the 96-point power prediction method to infer the 96-point power curve of the day to be predicted using the 96-point power prediction model.

[0316] Specifically, continue to refer to Figure 12b As shown in the figure, based on the pre-processed meteorological forecast data of the day to be predicted and the clear sky standard curve of the day to be predicted, the input features are constructed for each of the 96 points according to the feature construction method in S06, including the conventional meteorological features, timestamp features, power hysteresis features, harmonic irradiance features considering conversion efficiency, clear sky standard curve features, etc. corresponding to each point. The constructed feature data is input into the trained LightGBM model for inference to predict the 96-point power prediction curve within one day of the day to be predicted at one time.

[0317] S16. Construction of prediction interval based on daily power maximum prediction and standard curve. Although the prediction model for cloudy and rainy weather described in S07 has added the clear sky standard curve feature, the influence of this feature may be weakened during model learning. Therefore, it is necessary to perform forced post-processing based on this type of curve feature again to adjust the prediction points that do not conform to the regular cognition of the curve. Among them, the power curve prediction method under the sunny day classification described in S14 can also be directly applied to cloudy and rainy weather, but unlike sunny days with higher certainty, it is necessary to introduce a certain degree of fault tolerance and consider the situation of weather changes during the day (such as sunny and cloudy).

[0318] Continue to refer Fig.12c As shown, after obtaining the 96-point power curve within one day of the day to be predicted through power prediction under cloudy and rainy weather classification, the prediction interval construction (or post-processing interval construction) is performed. Specifically, in the reasoning stage described in S13, the weather type field of the meteorological forecast data of the day to be predicted is set to three preset weather types of "sunny", "cloudy", and "rainy and extreme weather" respectively, and the daily maximum power prediction model is used for re-reasoning to obtain the daily maximum power value prediction results under the three weather types.

[0319] Then, continue to refer to Figure 12b As shown, based on the preprocessed meteorological forecast data, the clear sky standard curve of the day to be predicted, and the daily maximum power prediction model, a post-processing interval is constructed for the daily maximum power value prediction results under all preset weather types.

[0320] Specifically, for each set weather type, the corresponding daily maximum power value prediction result is multiplied by each value in the clear sky standard curve of the day to be predicted to obtain the standard prediction curve under each set weather type. According to the weather type prediction for each time period of the day to be predicted in the meteorological forecast data, smoothing is performed (to prevent weather type jitter), and then the above three standard prediction curves are spliced ​​according to the corresponding time period (such as the morning period is sunny, the morning period adopts the standard prediction curve of the sunny type obtained by reasoning, and the afternoon is cloudy for two hours, then the corresponding standard curve of the cloudy type obtained by reasoning is adopted), so as to obtain the standard curve of the day to be predicted, so as to perform interval post-processing on the power prediction curve obtained in S15 based on the standard curve.

[0321] For the description of introducing the error tolerance percentage into the standard curve to calculate the confidence interval, please refer to the description of error tolerance correction data in this manual, and the details will not be repeated here.

[0322] S17, post-process the prediction results based on the prediction interval. For each point in the 96-point prediction curve of the day to be predicted under the cloudy or rainy weather type obtained in S15, the prediction result of the original LightGBM model is corrected according to the confidence interval at the corresponding moment. For a specific description, please refer to the description of correcting the first power prediction result according to the fault-tolerant correction data in this manual, which will not be repeated here.

[0323] Furthermore, the corrected prediction curve may have a sudden change in power value, so continue to refer to Figure 12b As shown in FIG. 1 , after the 96-point power prediction curve of the prediction day is post-processed by the post-processing interval, the modified prediction curve can be post-processed by a filter (e.g., a Savitzky-Golay filter, etc.) to smooth the curve, and then the final 96-point power prediction curve is output. Thus, forced post-processing based on the characteristics of the clear sky standard curve is achieved to adjust the prediction points that do not conform to the cognition of the power change law.

[0324] The above-mentioned photovoltaic power generation short-term power prediction method considering the clear sky law starts from the perspective of the main influencing factors of photovoltaic power generation, that is, the change law of photovoltaic power generation is mainly affected by the change law of solar irradiance, and the main influencing factors of solar irradiance are summarized as the relative position law of the earth and the sun under clear sky conditions and the influence of atmospheric clouds. Starting from the analysis of historical data, based on the historical data of photovoltaic power generation, the power curve under clear sky conditions is analyzed, and the concept of clear sky standard curve is proposed to explore the potential fixed laws brought by astronomical factors and enable the power prediction model to better learn the characteristics of this law.

[0325] Furthermore, for non-clear sky weather types such as cloudy and rainy, the post-processing logic is designed based on the above regular characteristics, and the prediction interval is constructed based on the daily maximum power value prediction model and the clear sky standard curve. The original power prediction results under non-clear sky weather are post-processed to impose constraints on the original prediction results using the clear sky law. In this way, the extracted prior knowledge (i.e., the clear sky standard curve) can be used to further improve the reliability of the prediction results of the automated algorithm.

[0326] By studying the clear sky law and the extracted clear sky standard curve characteristics, the prediction accuracy of the prediction model can be greatly improved. Taking a centralized photovoltaic power station as an example, for power prediction under sunny weather types, the power prediction accuracy of the power prediction method in the relevant technology is 83.6%. After using the power prediction method proposed in this specification, the prediction accuracy is increased to 92.3%. According to the grid-connected management standards specified by the provincial dispatch, the compliance rate on sunny days reaches more than 98%, and the model is more sensitive to the changes in the curve law caused by the switching of different seasons and months in a year. Compared with the prediction method in the relevant technology, the adjustment is more timely and can better fit the actual power generation curve shape. For the prediction scenarios under cloudy and rainy weather types, the short-term prediction accuracy is increased from 52.3% to 61.8%, which overall increases the number of days that the power forecast assessment of the power station meets the standards and reduces the power assessment of the power station.

[0327] The present specification also provides a data sample construction device, which may include:

[0328] The first data acquisition module is used to obtain the first power benchmark data of a cloudy day; wherein the first power benchmark data is obtained based on the power data of a cloudy day adjacent to the cloudy day; the first power benchmark data is used to describe the regularity of the photovoltaic power generation power obtained under the weather conditions without cloud cover at the astronomical position of the earth and the sun corresponding to the cloudy day.

[0329] The second data acquisition module is used to acquire the non-clear sky meteorological data of the non-clear sky day and the power data of the first historical related day; wherein the first historical related day is earlier than the non-clear sky day.

[0330] The feature construction module is used to construct feature values ​​based on the weather type corresponding to the non-clear sky day, the non-clear sky meteorological data, the power data of the first historical related day and the first power benchmark data to obtain the first non-clear sky sample for training the first initial prediction model.

[0331] The embodiments of this specification also provide a power prediction model training device, which may include:

[0332] The sample acquisition module is used to obtain non-clear sky power data on a non-clear sky day, and a first non-clear sky sample constructed by any data sample construction method in the aforementioned implementation manner.

[0333] The power prediction module is used to input the first non-clear sky sample into the first initial prediction model for prediction to obtain first power prediction data.

[0334] The model training module is used to train the first initial prediction model according to the first power prediction data and the non-clear sky power data to obtain the trained first power prediction model.

[0335] The embodiments of this specification also provide a power prediction device, which may include:

[0336] The third data acquisition module is used to obtain the meteorological forecast data of the day to be predicted and the second power benchmark data of the historical same period day; wherein the meteorological forecast data includes the forecast weather type corresponding to the day to be predicted; the forecast weather type includes the forecast whole day weather type; the date of the historical same period day and the date of the day to be predicted differ by an integer number of years; the second power benchmark data is used to describe the regularity of photovoltaic power generation power under the weather conditions without cloud cover at the astronomical position of the earth and the sun corresponding to the day to be predicted.

[0337] The historical power acquisition module is used to obtain the power data of the first historical day if the forecast weather type for the whole day is a non-clear sky type; wherein the first historical day is earlier than the day to be predicted.

[0338] The characteristic data construction module is used to construct the first weather characteristic data based on the forecast weather type, the meteorological forecast data, the power data of the first historical day and the second power reference data.

[0339] The prediction result acquisition module is used to input the first weather characteristic data into the first power prediction model trained by any power prediction model training method in the aforementioned implementation manner to obtain the first power prediction result of the day to be predicted.

[0340] For specific limitations on the data sample construction device, please refer to the limitations on the data sample construction method above, for specific limitations on the power prediction model training device, please refer to the limitations on the power prediction model training method above, and for specific limitations on the power prediction device, please refer to the limitations on the power prediction method above, which will not be repeated here. Each module in the above-mentioned data sample construction device, power prediction model training device, and power prediction device can be implemented in whole or in part through software, hardware, and a combination thereof. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.

[0341] An embodiment of the present specification also provides a computer device, which may include a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the data sample construction method, power prediction model training method, or power prediction method in any of the aforementioned embodiments.

[0342] The present specification also provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the data sample construction method or the power prediction model training method or the power prediction method in any of the aforementioned embodiments is implemented.

[0343] It should be noted that the logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be specifically implemented in any computer-readable medium for use by an instruction execution system, device or apparatus (such as a computer-based system, a system including a processor, or other system that can fetch instructions from an instruction execution system, device or apparatus and execute instructions), or in combination with these instruction execution systems, devices or apparatuses. For the purposes of this specification, "computer-readable medium" can be any device that can contain, store, communicate, propagate or transmit a program for use by an instruction execution system, device or apparatus, or in combination with these instruction execution systems, devices or apparatuses. More specific examples of computer-readable media (a non-exhaustive list) include the following: an electrical connection portion with one or more wirings (electronic device), a portable computer disk box (magnetic device), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable compact disk read-only memory (CDROM). In addition, the computer-readable medium may even be paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium and then editing, interpreting or processing in other suitable ways if necessary, and then stored in a computer memory.

[0344] It should be understood that the various parts of the present invention can be implemented by hardware, software, firmware or a combination thereof. In the above-mentioned embodiments, a plurality of steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, it can be implemented by any one of the following technologies known in the art or their combination: a discrete logic circuit having a logic gate circuit for implementing a logic function for a data signal, a dedicated integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0345] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "examples", "specific examples", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner.

[0346] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include at least one of the features. In the description of the present invention, the meaning of "plurality" is at least two, such as two, three, etc., unless otherwise clearly and specifically defined.

[0347] In the present invention, unless otherwise clearly specified and limited, the terms "installed", "connected", "connected", "fixed" and the like should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium, it can be the internal connection of two elements or the interaction relationship between two elements, unless otherwise clearly defined. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0348] Although the embodiments of the present invention have been shown and described above, it is to be understood that the above embodiments are exemplary and are not to be construed as limitations of the present invention. A person skilled in the art may change, modify, replace and vary the above embodiments within the scope of the present invention.

Claims

1. A data sample construction method, characterized in that: The method comprises: Acquire first power benchmark data for a cloudy day; wherein the first power benchmark data is obtained based on power data of a cloudy day adjacent to the cloudy day; the first power benchmark data is used to describe the regularity of the photovoltaic power generation power obtained under the weather conditions without cloud cover at the astronomical position of the earth and the sun corresponding to the cloudy day; Acquire the cloudy weather data of the cloudy day and the power data of the first historical related day; wherein the first historical related day is earlier than the cloudy day; Based on the weather type corresponding to the non-clear sky day, the non-clear sky meteorological data, the power data of the first historical related day and the first power benchmark data, feature values ​​are constructed to obtain a first non-clear sky sample for training a first initial prediction model.

2. The method according to claim 1, characterized in that The first power reference data is obtained according to power data of a clear sky day adjacent to the non-clear sky day, including: The first power reference data is obtained by performing interpolation calculation based on the power data of the first clear sky day and the power data of the second clear sky day adjacent to the clear sky day; wherein the first clear sky day is earlier than the clear sky day, and the second clear sky day is later than the clear sky day.

3. The method according to claim 1, characterized in that The non-clear sky day includes multiple collection time points; the weather type corresponding to the non-clear sky day includes the weather type corresponding to any collection time point; the non-clear sky meteorological data includes the first irradiance data and the first temperature data corresponding to any collection time point; the feature value construction based on the weather type corresponding to the non-clear sky day, the non-clear sky meteorological data, the power data of the first historical related day and the first power reference data includes: Acquiring preset operating temperature data of the photovoltaic power generation component and first operating temperature data at any of the acquisition time points; Determine the first harmonized irradiation data corresponding to any collection time point based on the preset working temperature and the first irradiation data, the first air temperature data and the first working temperature data corresponding to any collection time point; According to the weather types corresponding to each of the multiple collection time points, the non-clear sky meteorological data, the power data of the first historical related day, the first power benchmark data and the first harmonic irradiance data corresponding to each of the multiple collection time points, characteristic values ​​are constructed to obtain the first non-clear sky sample.

4. The method according to claim 1, characterized in that: The method further comprises: Obtaining clear sky meteorological data of a third clear sky day and power data of a second historical related day; wherein the weather type corresponding to the third clear sky day is a clear sky type; and the second historical related day is earlier than the clear sky day; Based on the clear sky type, the clear sky meteorological data and the power data of the second historical related day, feature values ​​are constructed to obtain clear sky samples for training a second initial prediction model.

5. The method according to claim 4, characterized in that The clear sky meteorological data includes second irradiance data and second temperature data corresponding to the maximum value in the clear sky power data of the third clear sky day; the characteristic value construction based on the clear sky type, the clear sky meteorological data and the power data of the second historical related day includes: Acquire preset operating temperature data of the photovoltaic power generation component and second operating temperature data corresponding to the maximum value in the clear sky power data; determining second blended irradiance data based on the second irradiance data, the second air temperature data, the preset working temperature data, and the second working temperature data; Characteristic values ​​are constructed according to the clear sky type, the clear sky meteorological data, the power data of the second historical related day, and the second harmonic irradiance data to obtain the clear sky sample.

6. The method according to claim 1, characterized in that The weather type corresponding to the non-clear sky day includes the day weather type of the non-clear sky day; the method further includes: Acquire power data of a third historical related day; wherein the third historical related day is earlier than the non-clear sky day; Based on the daily weather type of the non-clear sky day, the non-clear sky meteorological data and the power data of the third historical related day, feature values ​​are constructed to obtain a second non-clear sky sample for training a third initial prediction model.

7. The method according to claim 6, characterized in that The non-clear sky day corresponds to non-clear sky power data; the non-clear sky meteorological data includes third irradiance data and third temperature data corresponding to the maximum value in the non-clear sky power data; the characteristic value construction based on the daily weather type of the non-clear sky day, the non-clear sky meteorological data and the power data of the third historical related day includes: Acquire preset operating temperature data of the photovoltaic power generation component and third operating temperature data corresponding to the maximum value in the non-clear sky power data; Determine third blended irradiance data based on the third irradiance data, the third air temperature data, the preset working temperature data and the third working temperature data; Characteristic values ​​are constructed according to the daily weather type of the non-clear sky day, the non-clear sky meteorological data, the power data of the third historical related day and the third harmonic irradiance data to obtain the second non-clear sky sample.

8. A power prediction model training method, characterized in that: The method comprises: Acquire non-clear sky power data of a non-clear sky day, and a first non-clear sky sample constructed by the data sample construction method according to any one of claims 1 to 7; Inputting the first non-clear sky sample into a first initial prediction model for prediction to obtain first power prediction data; The first initial prediction model is trained according to the first power prediction data and the non-clear sky power data to obtain a trained first power prediction model.

9. The method according to claim 8, characterized in that The method further comprises: Obtain clear sky power data on a clear sky day, and a clear sky sample constructed by the data sample construction method according to any one of claims 4 to 5; Inputting the clear sky sample into a second initial prediction model for prediction to obtain second power prediction data; The second initial prediction model is trained according to the second power prediction data and the clear sky power data to obtain a trained second power prediction model.

10. The method according to claim 9, characterized in that The training of the second initial prediction model according to the second power prediction data and the clear sky power data comprises: The second initial prediction model is trained according to the maximum value of the second power prediction data and the clear sky power data to obtain the second power prediction model.

11. The method according to claim 9, characterized in that The second initial prediction model includes a first sub-model and a second sub-model; the step of inputting the clear sky sample into the second initial prediction model for prediction to obtain second power prediction data includes: Inputting the clear sky sample into the first sub-model for prediction to obtain power prediction data output by the first sub-model; Input the clear sky sample into the second sub-model for prediction, and obtain power prediction data output by the second sub-model; wherein the second power prediction data includes the power prediction data output by the first sub-model and the power prediction data output by the second sub-model; The step of training the second initial prediction model according to the second power prediction data and the clear sky power data to obtain a trained second power prediction model includes: The first sub-model is trained according to the power prediction data output by the first sub-model and the clear sky power data to obtain a first trained sub-model; The second sub-model is trained according to the power prediction data output by the second sub-model and the clear sky power data to obtain a second trained sub-model; wherein the second power prediction model includes the first trained sub-model and the second trained sub-model.

12. The method according to claim 8, characterized in that The method further comprises: Obtain a second non-clear sky sample constructed by the data sample construction method according to any one of claims 6 to 7; Inputting the second non-clear sky sample into a third initial prediction model for prediction to obtain third power prediction data; The third initial prediction model is trained according to the third power prediction data and the non-clear sky power data to obtain a trained third power prediction model.

13. The method according to claim 12, characterized in that The third power prediction model and the second power prediction model are the same power prediction model; wherein, the second power prediction model is obtained by training the second initial prediction model based on the second power prediction data of a clear sky day and the clear sky power data; the second power prediction data is obtained by predicting the second initial prediction model based on the input clear sky sample of the clear sky day.

14. A power prediction method, characterized in that: The method comprises: Obtaining meteorological forecast data for the day to be predicted and second power benchmark data for the same historical day; wherein the meteorological forecast data includes the forecast weather type corresponding to the day to be predicted; the forecast weather type includes the forecast weather type for the entire day; the date of the same historical day differs from the date of the day to be predicted by an integer number of years; the second power benchmark data is used to describe the regularity of photovoltaic power generation under the weather conditions without cloud cover at the astronomical position of the earth and the sun corresponding to the day to be predicted; If the forecast weather type for the whole day is a non-clear sky type, obtaining power data of the first historical day; wherein the first historical day is earlier than the day to be predicted; constructing first weather characteristic data based on the forecast weather type, the meteorological forecast data, the power data of the first historical day, and the second power reference data; The first weather characteristic data is input into a first power prediction model trained by any one of the methods of claims 8 to 13 to obtain a first power prediction result for the day to be predicted.

15. The method according to claim 14, characterized in that The method further comprises: If the forecast weather type for the whole day is a clear sky type, obtaining power data of a second historical day; wherein the second historical day is earlier than the day to be predicted; constructing second weather characteristic data based on the clear sky type, the weather forecast data and the power data of the second historical day; The second weather characteristic data is input into a second power prediction model trained by the method of any one of claims 9 to 11 to obtain a second power prediction result for the day to be predicted.

16. The method according to claim 15, characterized in that The step of inputting the second weather characteristic data into a second power prediction model trained by the method of any one of claims 9 to 11 to obtain a second power prediction result for the day to be predicted comprises: Inputting the second weather characteristic data into the second power prediction model to obtain a clear sky power prediction result output by the power prediction model; wherein the clear sky power prediction result is the predicted maximum power data of the day to be predicted; Numerical calculation is performed based on the clear sky power prediction result and the second power reference data to obtain the second power prediction result.

17. The method according to claim 14, characterized in that If the forecast weather type for the whole day is a non-clear sky type, the method further includes: Acquire power data of a third historical day, and determine a preset weather type for the day to be predicted; wherein the third historical day is earlier than the day to be predicted; constructing third weather characteristic data based on the preset weather type, the meteorological forecast data and the power data of the third historical day; Inputting the third weather characteristic data into the second power prediction model trained by the method of claim 13 to obtain a power prediction reference result corresponding to the preset weather type; The first power prediction result is corrected based on the power prediction reference result to obtain a corrected power prediction result for the day to be predicted.

18. The method according to claim 17, characterized in that The forecast weather type includes the predicted weather type of any period of the day to be predicted; the first power prediction result is corrected based on the power prediction reference result to obtain the corrected power prediction result of the day to be predicted, including: Determine, according to the power prediction reference result, the time period power reference data corresponding to the weather type of any time period; According to the time sequence of the multiple time periods of the day to be predicted, the time period power reference data corresponding to each of the multiple time periods are spliced ​​to obtain spliced ​​power data; Calculate according to the spliced ​​power data and preset fault tolerance data to obtain fault tolerance correction data; The first power prediction result is corrected according to the fault-tolerant correction data to obtain the corrected power prediction result.

19. The method according to claim 18, characterized in that The to-be-predicted day includes a plurality of to-be-predicted time points; the first power prediction result includes power data at any to-be-predicted time point; the fault-tolerant correction data includes correction data at any to-be-predicted time point; The correcting the first power prediction result according to the fault-tolerant correction data includes: If the power data of any of the to-be-predicted time points is within the range of the correction data of the corresponding to-be-predicted time point, the power data of the corresponding to-be-predicted time point is kept unchanged; If the power data of any of the to-be-predicted time points is greater than the maximum value of the corrected data of the corresponding to-be-predicted time point, the power data of the corresponding to-be-predicted time point is corrected to the maximum value of the corrected data of the corresponding to-be-predicted time point; If the power data at any of the to-be-predicted time points is less than the corrected data at the corresponding to-be-predicted time points, the power data at the corresponding to-be-predicted time points is corrected to the minimum value of the corrected data at the corresponding to-be-predicted time points.

20. A data sample construction device, characterized in that: The device comprises: A first data acquisition module is used to acquire first power benchmark data of a cloudy day; wherein the first power benchmark data is obtained based on power data of a cloudy day adjacent to the cloudy day; the first power benchmark data is used to describe the regularity of the photovoltaic power generation power obtained under the weather conditions without cloud cover at the astronomical position of the earth and the sun corresponding to the cloudy day; A second data acquisition module is used to acquire the non-clear sky meteorological data of the non-clear sky day and the power data of the first historical related day; wherein the first historical related day is earlier than the non-clear sky day; A feature construction module is used to construct feature values ​​based on the weather type corresponding to the non-clear sky day, the non-clear sky meteorological data, the power data of the first historical related day and the first power benchmark data to obtain a first non-clear sky sample for training a first initial prediction model.

21. A power prediction model training device, characterized in that: The device comprises: A sample acquisition module, used to acquire non-clear sky power data on a non-clear sky day, and a first non-clear sky sample constructed by the data sample construction method according to any one of claims 1 to 7; A power prediction module, used for inputting the first non-clear sky sample into a first initial prediction model for prediction to obtain first power prediction data; A model training module is used to train the first initial prediction model according to the first power prediction data and the non-clear sky power data to obtain a trained first power prediction model.

22. A power prediction device, characterized in that: The device comprises: The third data acquisition module is used to acquire the meteorological forecast data of the day to be predicted and the second power benchmark data of the historical day of the same period; wherein the meteorological forecast data includes the forecast weather type corresponding to the day to be predicted; the forecast weather type includes the forecast weather type for the whole day; the date of the historical day of the same period differs from the date of the day to be predicted by an integer number of years; the second power benchmark data is used to describe the regularity of photovoltaic power generation under the weather conditions without cloud cover at the astronomical position of the earth and the sun corresponding to the day to be predicted; A historical power acquisition module, used for acquiring power data of a first historical day if the forecast weather type for the whole day is a non-clear sky type; wherein the first historical day is earlier than the day to be predicted; A characteristic data construction module, configured to construct first weather characteristic data based on the forecast weather type, the meteorological forecast data, the power data of the first historical day and the second power reference data; A prediction result acquisition module is used to input the first weather characteristic data into a first power prediction model trained by any one of the methods of claims 8 to 13 to obtain a first power prediction result for the day to be predicted.

23. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 19 are implemented.

24. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 19 are implemented.

Images