Photovoltaic optical power prediction method and system based on large model, and medium

Abstract

The invention discloses a photovoltaic optical power prediction method and system based on a large model, and a medium, mainly relates to the technical field of photovoltaic optical power prediction, and is used for solving the problems of high model training cost and significant data island problem of an existing scheme. Comprising the following steps: collecting optical power associated data, and preprocessing optical power and the optical power associated data to obtain a training data set; modeling the pre-trained large model of the time sequence to obtain an initial large model; injecting a trainable parameter matrix into the pre-training model to obtain a large model for training; obtaining a prompt template, inputting the training data set into the prompt template to obtain prompt training data, and inputting the prompt training data into the large model for training to obtain a trained large model; and acquiring real-time optical power associated data, and inputting the real-time optical power associated data into the trained large model to obtain predicted optical power.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of photovoltaic power prediction, and in particular to a photovoltaic power prediction method and system based on a large model and a medium. BACKGROUND

[0002] The current mainstream photovoltaic power prediction algorithm mainly includes three types: physical model method, statistical model method and artificial intelligence model method. The physical model method is based on the physical principle of photovoltaic power generation, and a mathematical model is established to describe the light radiation transmission and the photovoltaic component photoelectric conversion process. The statistical model method is based on the statistical law of historical data, and a mapping relationship between input variables (such as time, meteorological parameters) and output power is established. The artificial intelligence model method automatically extracts data features through a machine learning algorithm to capture complex nonlinear relationships.

[0003] The above-mentioned scheme has the following problems: under the background of rapid expansion of the photovoltaic industry, when a large energy enterprise manages hundreds to thousands of stations, the traditional method needs to invest independent modeling manpower, computing power and time cost for each station. Station equipment modification (such as replacing inverters, cleaning components) or environmental changes (such as new high-rise buildings built around) will cause data distribution drift, and thus lead to high model training cost. In addition, the data of the station is not effectively integrated and utilized, and the historical data cannot be mutually assisted in modeling, which wastes the value of data. SUMMARY

[0004] The present application provides a photovoltaic power prediction method and system based on a large model to solve the problems of high model training cost and significant data island in the prior art.

[0005] In a first aspect, the present application provides a photovoltaic power prediction method based on a large model, which comprises: Collecting light power related data, preprocessing the light power and the light power related data to obtain a training data set; wherein the light power related data at least includes: meteorological data, photovoltaic power station operation data, geographic information data, and time information data; Modeling a pre-trained large model of time series to obtain an initial large model; injecting a trainable parameter matrix into the pre-trained model to obtain a large model for training; Obtaining a prompt template, inputting the training data set into the prompt template to obtain prompt training data, inputting the prompt training data into the large model for training to obtain a trained large model; Obtaining real-time light power related data, inputting the real-time light power related data into the trained large model to obtain predicted light power.

[0006] In an implementation manner of the present application, the light power related data is preprocessed to obtain a training data set, which specifically includes: combining the optical power and the optical power associated data with the same timestamp into an optical power associated field according to the optical power and the timestamp of the optical power associated data; statistically counting a missing rate of data in each optical power associated field, and directly deleting data with a missing rate greater than a preset missing threshold; obtaining an optical power associated field with missing time information data, and filling in the time information data by using time series interpolation or neighborhood padding; obtaining a preset normal data interval of the optical power associated data, and removing optical power associated fields not belonging to the preset normal data interval; obtaining seasonal data with seasonal changes in the optical power associated field; establishing a seasonal data-optical power curve, normalizing the seasonal data-optical power curve, and correcting the seasonal data; normalizing non-seasonal data in the optical power associated field to obtain a normalized optical power associated field; obtaining a set division rule, and dividing the power associated field into a plurality of data sets according to the set division rule; adding a preset amplitude Gaussian noise to the optical power in the data set; obtaining a preset scarcity scenario condition, and then obtaining a corresponding power associated field from the data set; synthesizing a simulation sample by using a generative adversarial network, and then adding the simulation sample to the corresponding data set to obtain a training data set.

[0007] In an implementation manner of the present application, the preset normal data interval of the optical power associated data specifically includes: calculating a mean value of first data conforming to a normal distribution in the optical power associated data, and determining the mean value the range of the preset correction value is the preset normal data interval of the first data; sorting second data conforming to a non-normal distribution in the optical power associated data from small to large, dividing the second data into four parts, determining a first data of the second part as Q1 and a first data of the fourth part as Q3, and determining [Q1-(Q3-Q1), Q3+(Q3-Q1)] as the preset normal data interval of the second data.

[0008] In an implementation manner of the present application, the set division rule is obtained, and the power associated field is divided into a plurality of data sets according to the set division rule, specifically including: dividing the power associated field by year or quarter; or, dividing the power associated field by climate zone grouping.

[0009] In an implementation manner of the present application, before the prompt training data is input into the large model for training to obtain the trained large model, the method further includes: The training parameters are configured; wherein the training parameters at least include: the learning rate is: 1e-5 to 1e-6; the training period is: 500-1000 steps.

[0010] In an implementation manner of the present application, the prompt template at least includes: predicted photovoltaic power station power, time, irradiance, temperature, historical power sequence, irradiance-power linear relationship and time trend, and predicted power value.

[0011] In a second aspect, the present application provides a photovoltaic power prediction system based on a large model, the system comprising: The obtaining module is configured to collect photovoltaic power correlation data, pre-process the photovoltaic power and the photovoltaic power correlation data, and obtain a training data set; wherein the photovoltaic power correlation data at least includes: meteorological data, photovoltaic power station operation data, geographic information data, and time information data; The model module is configured to model a pre-trained large model of time series, obtain an initial large model, inject a trainable parameter matrix into the pre-trained model, obtain a large model for training, input the training data set into the prompt template to obtain prompt training data, input the prompt training data into the large model for training, and obtain a trained large model; The prediction module is configured to obtain real-time photovoltaic power correlation data, input the real-time photovoltaic power correlation data into the trained large model, and obtain predicted photovoltaic power.

[0012] In an implementation manner of the present application, the obtaining module comprises an obtaining unit, configured to combine the photovoltaic power and the photovoltaic power correlation data of the same timestamp into a set of photovoltaic power correlation fields based on the timestamp on the photovoltaic power and the photovoltaic power correlation data; configured to statistically calculate the missing rate of data in each photovoltaic power correlation field, and directly delete data with a missing rate greater than a preset missing threshold; configured to obtain photovoltaic power correlation fields with missing time information data, and fill in the time information data by using time series interpolation or neighborhood padding; configured to obtain a preset normal data interval of the photovoltaic power correlation data, and remove photovoltaic power correlation fields not belonging to the preset normal data interval; configured to obtain seasonal data with seasonal changes in the photovoltaic power correlation fields; configured to establish a seasonal data-photovoltaic power curve, normalize the seasonal data-photovoltaic power curve, and correct the seasonal data; configured to normalize the non-seasonal data in the photovoltaic power correlation fields to obtain normalized photovoltaic power correlation fields; configured to obtain a set division rule, and divide the power correlation fields into a plurality of data sets according to the set division rule; configured to add a preset amplitude Gaussian noise to the photovoltaic power in the data set; obtain a preset scarce scene condition, and then obtain a corresponding power correlation field from the data set; By generating a generative adversarial network, a simulation sample is synthesized, and then the simulation sample is added to the corresponding data set to obtain a training data set.

[0013] In an implementation form of the application, the obtaining unit comprises an obtaining subunit, for calculating the mean value of the first data conforming to the normal distribution in the optical power correlation data, determining the mean value The range of the preset correction value is a preset normal data interval of the first data. The second data conforming to the non-normal distribution in the optical power correlation data is sorted from small to large and divided into four parts, the first data of the second part is determined as Q1, and the first data of the fourth part is determined as Q3; and [Q1-(Q3-Q1), Q3+(Q3-Q1)] is determined as the preset normal data interval of the second data.

[0014] In a third aspect, the application provides a non-volatile computer storage medium having computer instructions stored thereon, the computer instructions being executed to implement a photovoltaic power prediction method based on a large model according to any one of the above.

[0015] As can be seen from the above technical solutions, the application has the following advantages: By constructing a unified pre-training large model framework, heterogeneous data such as meteorological data, operation data, and geographic information are integrated into a standardized training set, realizing the integration of multi-source data and avoiding the problem of data silos. The application injects a trainable parameter matrix into the pre-training model to obtain a large model for training, so that new stations only need to inject a differentiated parameter matrix to complete the adaptation, solving the problem of linear growth of training cost. In addition, the traditional method needs to retrain the model in full when the equipment is modified (such as replacing the inverter) or the environment changes (such as adding a shelter), while the application only needs to perform incremental update on the affected module through the local adjustment mechanism of the trainable parameter matrix. BRIEF DESCRIPTION OF DRAWINGS

[0016] In order to more clearly illustrate the technical solutions of the present application, the following will briefly introduce the drawings needed to be used in the description. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can also be obtained by those skilled in the art without creative labor.

[0017] Figure 1 is a flow chart of a photovoltaic power prediction method based on a large model provided by an embodiment of the present application.

[0018] Figure 2It is an internal structure schematic diagram of a photovoltaic light power prediction system based on a large model provided by the embodiment of the application. DETAILED DESCRIPTION

[0019] The technical solutions in the embodiments of the application will be clearly and completely described below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by a person of ordinary skill in the art without creative work are within the protection scope of the application.

[0020] It should be understood by those skilled in the art that the embodiments described below are only preferred embodiments of the present disclosure, and do not represent that the present disclosure can only be implemented by the preferred embodiments. The preferred embodiments are only used to explain the technical principles of the present disclosure, and are not used to limit the protection scope of the present disclosure. Based on the preferred embodiments provided by the present disclosure, all other embodiments obtained by a person of ordinary skill in the art without creative work are still within the protection scope of the present disclosure.

[0021] It should be further noted that the terms "comprise", "contain" or any other variants thereof are intended to cover non-exclusive inclusion, so that the process, method, product or equipment including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such process, method, product or equipment. Without more limitations, the element defined by the statement "comprises a" does not exclude the presence of another identical element in the process, method, product or equipment including the element.

[0022] The technical solutions provided by the embodiments of the application will be described in detail below with reference to the drawings.

[0023] The embodiment provides a photovoltaic light power prediction method based on a large model. Figure 1 As shown in the figure, the method provided by the embodiment of the application mainly includes the following steps: Step 110, collecting light power correlation data, pre-processing the light power and light power correlation data to obtain a training data set.

[0024] The light power correlation data at least includes meteorological data, photovoltaic power station operation data, geographic information data and time information data.

[0025] Further specifically, regarding real-time meteorological data: Real-time collection of solar radiation intensity, ambient temperature, humidity, wind speed, wind direction, atmospheric pressure data by installing meteorological monitoring equipment (such as solar radiation sensors, temperature sensors, humidity sensors, anemometers, barometers, etc.) on the photovoltaic power station site. At the same time, access to professional weather service provider data interfaces (such as weather bureau data services, AccuWeather, etc.) to obtain more comprehensive meteorological data, including cloud cover, precipitation probability, weather warning information, etc. Historical meteorological data: Obtain historical meteorological data of the target station area from meteorological databases, meteorological department public data platforms, or commercial data suppliers for model training and analysis of meteorological change patterns.

[0026] Photovoltaic power station operation data includes equipment operation parameters and system operation data.

[0027] Equipment operation parameters: Collect photovoltaic module working voltage, working current, output power, temperature, etc. through the monitoring system of the photovoltaic power station; Inverter input voltage, current, output voltage, current, power, conversion efficiency, working state (normal, fault, shutdown) and other parameters; Current, voltage, switch state and other parameters of the bus box. System operation data: Collect total power generation, on-grid power, abandoned light power, equipment fault records, maintenance records and other data of the power station, comprehensively reflect the operation state and performance of the power station.

[0028] Obtain the geographic location information of the photovoltaic power station, including accurate latitude and longitude coordinates, altitude, terrain (such as plains, mountains, hills, etc.), surrounding building distribution and height, etc. These data can be obtained through global positioning system (GPS), geographic information system (GIS) or field measurement. Geographic information data plays an important role in analyzing the shading and shadow changes of solar radiation.

[0029] Record the specific timestamp of data collection, including year, month, day, hour, minute, second, and corresponding season, week, holiday, etc. Time information is a key factor in analyzing the periodic changes of photovoltaic power (such as daily, weekly, seasonal changes).

[0030] In some embodiments, the light power related data is preprocessed to obtain a training data set, specifically including: Based on the timestamp on the light power and light power related data, the light power and light power related data of the same timestamp are combined into a set of light power related fields; Statistically analyze the missing rate of data in each light power related field, and directly delete data with a missing rate greater than a preset missing threshold (for example, 30%); Obtain the light power related field of the missing time information data, and fill in the time information data using time series interpolation or neighborhood padding; obtaining a preset normal data interval of the optical power correlation data, removing the optical power correlation field not belonging to the preset normal data interval; obtaining seasonal data existing in the optical power correlation field with seasonal change; establishing a seasonal data-optical power curve, normalizing the seasonal data-optical power curve, and correcting the seasonal data; normalizing the non-seasonal data in the optical power correlation field to obtain the normalized optical power correlation field; obtaining a set division rule, and dividing the power correlation field into a plurality of data sets according to the set division rule; adding a preset amplitude Gaussian noise to the optical power in the data set; obtaining a preset rare scene condition, and then obtaining the corresponding power correlation field from the data set; synthesizing a simulation sample through a generative adversarial network, and then adding the simulation sample to the corresponding data set to obtain a training data set.

[0031] The preset normal data interval of the optical power correlation data can be specifically obtained as follows: calculating the mean value of the first data conforming to the normal distribution in the optical power correlation data, and determining the mean value The range of the preset correction value is the preset normal data interval of the first data. The second data conforming to the non-normal distribution in the optical power correlation data is sorted from small to large, and divided into four parts, the first data of the second part is determined as Q1, and the first data of the fourth part is determined as Q3; and [Q1-(Q3-Q1), Q3+(Q3-Q1)] is determined as the preset normal data interval of the second data.

[0032] The set division rule can be specifically obtained as follows: The power correlation field is divided according to years or seasons; or, the power correlation field is divided according to climate (such as temperate zone / tropical zone / plateau).

[0033] Based on the above description, those skilled in the art can understand that the step ensures the integrity of the input features through multi-dimensional data collaborative collection. Through the combination of field meteorological monitoring equipment (radiation sensor, temperature and humidity instrument, etc.) and professional meteorological service interface (such as AccuWeather), real-time / historical data full coverage of key meteorological parameters such as solar radiation, temperature and humidity is realized. At the same time, the synchronous collection of power plant operation data (component voltage and current, inverter efficiency, junction box state, etc.) and geographic information (latitude, longitude, altitude, terrain) fully depicts the influence mechanism of the "environment-equipment-space" three elements on the light power. For example, the height data of surrounding buildings can be directly used for shadow shielding analysis, and the inverter state parameters can timely reflect the interference of equipment failure on power generation. This structured data collection system solves the problem of single data dimension in traditional methods.

[0034] The intelligent preprocessing process improves data usability. Through timestamp alignment, missing value processing (deleting fields with a missing rate of >30% or interpolating filling), and outlier removal (based on normal / non-normal distribution interval detection), the error rate of the original data is reduced to below 3%. Seasonal data normalization processing (such as establishing a seasonal-power curve) effectively eliminates data bias caused by climate periodic fluctuations, while Gaussian noise injection and generated adversarial network (GAN) simulation sample synthesis enhance the anti-interference ability and coverage rate of rare scenarios of the data set. The design of dividing the data set according to climate zones or time periods makes the training data distribution consistent with the actual physical law.

[0035] Step 120, modeling the pre-trained large model of the time series to obtain an initial large model; injecting a trainable parameter matrix into the pre-trained model to obtain a large model for training; obtaining a prompt template, inputting the training data set into the prompt template to obtain prompt training data, and inputting the prompt training data into the large model for training to obtain a trained large model.

[0036] It should be noted that the pre-trained large model can be DeepSeek-TS.

[0037] Before inputting the prompt training data into the large model for training to obtain the trained large model, the method further comprises: configuring training parameters; wherein the training parameters at least include: the learning rate is between 1e-5 and 1e-6; the training period is between 500-1000 steps.

[0038] The prompt template at least includes: predicted photovoltaic power station power, time, irradiance, temperature, historical power sequence, irradiance-power linear relationship and time series trend, and predicted power value.

[0039] It should be noted that, since the application injects a trainable parameter matrix, only tens to hundreds of samples (such as historical meteorological data and power data of a single station) are needed in the training process to complete the adaptation, solving the problem of insufficient data accumulation of photovoltaic stations (especially new grid-connected stations and small distributed power stations).

[0040] Based on the above description, those skilled in the art can understand that this step improves the model generalization ability by pre-training the large model architecture. Specifically, a time series pre-training model such as DeepSeek-TS is used as the basic architecture, which has learned the general feature representation rules through massive time series data. By injecting a trainable parameter matrix (about 1%-5% of the total model parameters), only a small number of parameters need to be adjusted to adapt to the characteristics of a specific station while preserving the pre-training knowledge. This "frozen backbone + fine-tuning adaptation layer" design allows the model to converge within 500-1000 training steps (learning rate 1e-5 to 1e-6), which is shorter than the training period of traditional models. At the same time, since the pre-training model already has time series feature extraction capability, only tens of historical data are needed for new stations to complete the adaptation, solving the pain point of insufficient data for small-scale power stations.

[0041] In addition, this step converts heterogeneous information such as meteorological data and operating parameters into context descriptions understandable by the model through structured prompt templates (including irradiance-power relationship, historical sequence, etc.). For example, the temperature sensor reading and the corresponding historical power sequence with time stamp are jointly encoded as a "temperature-power" time series prompt, allowing the model to explicitly capture the lagging effect of temperature on power output. This template-based input allows the model to automatically associate linear relationships between irradiance, device status, and other multi-dimensional features during prediction, reducing manual intervention compared to traditional feature engineering methods.

[0042] Step 130, obtaining real-time light power correlation data, inputting the real-time light power correlation data into the trained large model to obtain the predicted light power.

[0043] Based on the foregoing description, the present embodiment solves the problem of single data dimension in traditional methods by constructing a complete data set containing meteorological, device, geographical, and time dimensions through multi-source data fusion (on-site monitoring equipment + professional meteorological interface + power station operation system). For example, the cloud coverage in real-time meteorological data combined with the surrounding building height in the geographical information of the power station can directly quantify the impact of shadow shielding on power; and the correlation analysis of inverter efficiency parameters and historical power sequence can accurately capture the power generation efficiency decay caused by device aging. The preprocessing step performs timestamp alignment, outlier removal (based on normal / tetrad interval detection), and other operations, and normalizes seasonal data to eliminate the interference of climate cycle fluctuations on model training, ensuring the accuracy and consistency of the input data.

[0044] The DeepSeek-TS pre-training large model is adopted, a trainable parameter matrix (only accounting for 1%-5% of the total parameters of the model) is injected to realize "main body freezing + local fine-tuning", and the specific station adaptation can be completed within 500-1000 training cycles, which improves the training efficiency compared with the traditional method. Especially for new grid-connected stations or distributed small power stations, only dozens of historical data are needed to meet the prediction accuracy requirements, solving the pain point of insufficient data accumulation. The structured prompt template converts multi-dimensional data into the context description of "irradiance-power relationship + time trend", so that the model can automatically associate complex characteristics such as temperature lag effect and equipment failure state, and reduces manual intervention compared with traditional feature engineering.

[0045] In addition, the application Figure 2 A photovoltaic power prediction system based on a large model is provided for an embodiment of the application. As Figure 2 shown, the system provided by the embodiment of the application mainly comprises: The obtaining module 210 is configured to collect photovoltaic power correlation data, pre-process the photovoltaic power and the photovoltaic power correlation data, and obtain a training data set; wherein the photovoltaic power correlation data at least includes meteorological data, photovoltaic power station operation data, geographic information data, and time information data.

[0046] The obtaining module 210 comprises an obtaining unit, configured to combine the photovoltaic power and the photovoltaic power correlation data of the same timestamp as a group of photovoltaic power correlation fields based on the timestamp on the photovoltaic power and the photovoltaic power correlation data; configured to directly delete the data with a missing rate greater than a preset missing threshold; configured to obtain the photovoltaic power correlation fields with missing time information data, and fill in the time information data by using time series interpolation or neighborhood padding; configured to obtain a preset normal data interval of the photovoltaic power correlation data, and remove the photovoltaic power correlation fields not belonging to the preset normal data interval; configured to obtain seasonal data with seasonal changes in the photovoltaic power correlation fields; configured to establish a seasonal data-photovoltaic power curve, normalize the seasonal data-photovoltaic power curve, and correct the seasonal data; configured to normalize the non-seasonal data in the photovoltaic power correlation fields to obtain normalized photovoltaic power correlation fields; configured to obtain a set division rule, and divide the power correlation fields into a plurality of data sets according to the set division rule; configured to add a preset amplitude Gaussian noise to the photovoltaic power in the data set; configured to obtain a preset scarce scenario condition, and then obtain the corresponding power correlation fields from the data set; The simulation sample is synthesized through the generative adversarial network, and then the simulation sample is added to the corresponding data set to obtain the training data set.

[0047] The obtaining unit comprises an obtaining subunit, The mean value of the first data conforming to the normal distribution in the optical power correlation data is calculated, and the mean value The range of the preset correction value is a preset normal data interval of the first data. The second data conforming to the non-normal distribution in the optical power correlation data is sorted from small to large and divided into four parts, the first data of the second part is Q1, and the first data of the fourth part is Q3; and [Q1-(Q3-Q1), Q3+(Q3-Q1)] is determined as the preset normal data interval of the second data.

[0048] The model module 220 is configured to model the pre-trained large model of the time series, obtain an initial large model, inject a trainable parameter matrix into the pre-trained model, obtain a large model for training, obtain a prompt template, input the training data set into the prompt template, obtain prompt training data, and input the prompt training data into the large model for training to obtain a trained large model.

[0049] The prediction module 230 is configured to obtain real-time optical power correlation data, input the real-time optical power correlation data into the trained large model, and obtain predicted optical power.

[0050] In addition, the embodiment of the present application further provides a non-volatile computer storage medium having executable instructions stored thereon, which, when executed, implement a large model-based photovoltaic optical power prediction method as described above.

[0051] The above description of the disclosed embodiments enables a person skilled in the art to implement or use the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A photovoltaic light power prediction method based on a large model, characterized in that: The method comprises: Collecting optical power related data, pre-processing the optical power and optical power related data to obtain a training data set; wherein the optical power related data at least includes: meteorological data, photovoltaic power station operation data, geographic information data, and time information data; Modeling the pre-trained large model of the time series to obtain an initial large model; injecting a trainable parameter matrix into the pre-trained model to obtain a training large model; obtaining a prompt template, inputting the training data set into the prompt template to obtain prompt training data, and inputting the prompt training data into the training large model to obtain a trained large model; Real-time optical power correlation data is obtained and input into the trained large model to obtain predicted optical power.

2. The photovoltaic light power prediction method based on a large model according to claim 1, characterized in that: Preprocess the optical power related data to obtain a training data set, including: Based on the timestamps on the optical power and the optical power-related data, the optical power and the optical power-related data with the same timestamp are combined into a group of optical power-related fields; Count the missing rate of data in each optical power-related field, and directly delete the data whose missing rate is greater than the preset missing threshold; Obtain the optical power related fields of the missing time information data and use time series interpolation or neighborhood filling to complete the time information data; Obtaining a preset normal data interval of optical power related data, and removing optical power related fields that do not belong to the preset normal data interval; Obtain seasonal data with seasonal variations in optical power related fields; Establish a seasonal data-optical power curve, normalize the seasonal data-optical power curve, and correct the seasonal data; Normalizing the non-seasonal data in the optical power correlation field to obtain a normalized optical power correlation field; Obtain a set partitioning rule, and divide the power-related field into several data sets according to the set partitioning rule; Adding Gaussian noise of preset amplitude to the optical power in the data set; Obtain the preset scarce scene conditions, and then obtain the corresponding power-related fields from the data set; By generating an adversarial network, we synthesize simulation samples, and then add the simulation samples to the corresponding data set to obtain a training data set.

3. The photovoltaic light power prediction method based on a large model according to claim 2, characterized in that: Obtain the preset normal data interval for optical power related data, specifically including: Calculate the mean of the first data that conforms to the normal distribution in the optical power correlation data and determine the mean The range of the preset correction value is a preset normal data interval of the first data; The second data in the optical power correlation data that conforms to the non-normal distribution is sorted from small to large and divided into four parts. The first data of the second part is determined to be Q1, and the first data of the fourth part is Q3; [Q1-(Q3-Q1), Q3+(Q3-Q1)] is determined to be the preset normal data interval of the second data.

4. The photovoltaic light power prediction method based on a large model according to claim 2, characterized in that: Obtain a set partitioning rule and divide the power-related fields into several data sets according to the set partitioning rule, including: Split power-related fields by year or quarter; Alternatively, group the power correlation fields by climate region.

5. The photovoltaic light power prediction method based on a large model according to claim 1, characterized in that: Before inputting the prompt training data into the training large model to obtain the trained large model, the method further includes: Configure training parameters; wherein the training parameters include at least: learning rate: 1e-5 to 1e-6; training cycle: 500-1000 steps.

6. The photovoltaic light power prediction method based on a large model according to claim 1, characterized in that: The prompt template includes at least: predicted photovoltaic station power, time, irradiance, temperature, historical power sequence, irradiance-power linear relationship and time series trend, and predicted power value.

7. A photovoltaic light power prediction system based on a large model, characterized in that: The system comprises: An acquisition module is used to collect optical power related data, pre-process the optical power and optical power related data, and obtain a training data set; wherein the optical power related data at least includes: meteorological data, photovoltaic power station operation data, geographic information data, and time information data; The model module is used to model the pre-trained large model of the time series and obtain the initial large model; inject the trainable parameter matrix into the pre-trained model to obtain the training large model; obtain the prompt template, input the training data set into the prompt template to obtain the prompt training data, input the prompt training data into the training large model to obtain the trained large model; The prediction module is used to obtain real-time optical power correlation data, input the real-time optical power correlation data into the trained large model, and obtain the predicted optical power.

8. The photovoltaic light power prediction system based on a large model according to claim 7, characterized in that: The acquisition module includes an acquisition unit, Based on the timestamps on the optical power and the optical power-related data, the optical power and the optical power-related data with the same timestamp are combined into a group of optical power-related fields; Count the missing rate of data in each optical power-related field, and directly delete the data whose missing rate is greater than the preset missing threshold; Obtain the optical power related fields of the missing time information data and use time series interpolation or neighborhood filling to complete the time information data; Obtaining a preset normal data interval of optical power related data, and removing optical power related fields that do not belong to the preset normal data interval; Obtain seasonal data with seasonal variations in optical power related fields; Establish a seasonal data-optical power curve, normalize the seasonal data-optical power curve, and correct the seasonal data; Normalizing the non-seasonal data in the optical power correlation field to obtain a normalized optical power correlation field; Obtain a set partitioning rule, and divide the power-related field into several data sets according to the set partitioning rule; Adding Gaussian noise of preset amplitude to the optical power in the data set; Obtain the preset scarce scene conditions, and then obtain the corresponding power-related fields from the data set; By generating an adversarial network, we synthesize simulation samples, and then add the simulation samples to the corresponding data set to obtain a training data set.

9. The photovoltaic light power prediction system based on a large model according to claim 8, characterized in that: The obtaining unit includes obtaining subunits, Used to calculate the mean of the first data that conforms to the normal distribution in the optical power correlation data, and determine the mean The range of the preset correction value is a preset normal data interval of the first data; The second data in the optical power correlation data that conforms to the non-normal distribution is sorted from small to large and divided into four parts. The first data of the second part is determined to be Q1, and the first data of the fourth part is Q3; [Q1-(Q3-Q1), Q3+(Q3-Q1)] is determined to be the preset normal data interval of the second data.

10. A non-volatile computer storage medium, characterized in that Computer instructions are stored thereon, and when the computer instructions are executed, the photovoltaic light power prediction method based on a large model as described in any one of claims 1 to 6 is implemented.

Images