Medium and short term photovoltaic output prediction method based on multi-day meteorological combination

Abstract

The invention provides a medium and short term photovoltaic output prediction method based on multi-day meteorological combination, comprising the following steps: obtaining feature data of a plurality of meteorological features in meteorological data corresponding to a prediction day, the meteorological data corresponding to the prediction day comprising meteorological data of three days before, on the same day and after the prediction day; extracting feature data of the important meteorological features based on a feature selector to form a feature set of a prediction day; similarities between the feature set and the meteorological clusters are calculated respectively, the output prediction model corresponding to the meteorological cluster with the highest similarity is called, and the output prediction model predicts an initial prediction value representing photovoltaic output based on the feature set of the prediction day. According to the method, when meteorological changes are complex or meteorological data are inaccurate, photovoltaic output in a short period can be efficiently and accurately predicted by combining multi-day meteorological data, the consumption efficiency of a power grid on the photovoltaic output is improved, the output of thermal power generation is reduced, and environmental pollution is reduced.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic power output prediction technology, specifically to a short-to-medium-term photovoltaic power output prediction method based on multi-day meteorological combinations. Background Technology

[0002] Thermal power generation relies on burning fossil fuels, which are finite and cause environmental pollution. Photovoltaic power generation, as a clean and renewable energy source, faces several key challenges in its widespread application. Photovoltaic power generation is affected by sunlight; its output varies with weather, season, and time, and its intermittent, random, and fluctuating nature can impact the power system. Therefore, accurate photovoltaic output forecasting is crucial for ensuring the reliability of the power system.

[0003] Existing methods for predicting photovoltaic (PV) output rely on cloud cover on the predicted day to estimate that output. However, these methods suffer from low accuracy when faced with complex weather patterns and inaccurate weather data. Complex weather patterns indicate frequent changes in weather within a single day, and inaccurate weather data is typically due to inaccurate daily weather forecasts from meteorological observatories.

[0004] When photovoltaic (PV) power output forecasts have significant deviations, it becomes impossible to fully integrate PV power into the grid. Therefore, a method is needed to improve the accuracy and stability of PV power output forecasts, thereby increasing the grid's efficiency in integrating PV power, reducing thermal power generation output, and lowering environmental pollution. Summary of the Invention

[0005] In view of this, the problem to be solved by the present invention is to provide a method for predicting short- and medium-term photovoltaic power output based on multi-day meteorological combinations, which can efficiently and accurately predict short-term photovoltaic power output when meteorological changes are complex or meteorological data is inaccurate, thereby improving the grid's absorption efficiency of photovoltaic power output, reducing thermal power generation output, and reducing environmental pollution.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A method for predicting short- and medium-term photovoltaic power output based on multi-day meteorological combinations includes acquiring feature data of several meteorological features within the meteorological data corresponding to the prediction day, and extracting feature data of important meteorological features based on a feature selector to form a feature set for the prediction day. The similarity between the feature set and several meteorological clusters is calculated respectively. The power output prediction model corresponding to the meteorological cluster with the highest similarity is retrieved. The power output prediction model predicts the initial predicted value representing the photovoltaic power output based on the feature set of the prediction day.

[0007] Furthermore, the meteorological data corresponding to the forecast date includes the actual meteorological data of the day before the forecast date, the meteorological forecast data of the forecast date, and the meteorological forecast data of the days following the forecast date.

[0008] Furthermore, the meteorological data includes temperature, humidity, wind speed, cloud cover, and solar irradiance; The meteorological characteristics include temperature, humidity, wind speed, cloud cover, and solar irradiance, with their respective mean, variance, maximum, and minimum values.

[0009] Furthermore, the determination of important meteorological features in the feature selector includes: acquiring feature data of several meteorological features and actual photovoltaic power output data from meteorological data corresponding to several historical days, and constructing several regression trees based on the random forest regression algorithm; calculating the amount of impurity reduction brought about by a certain meteorological feature at its split node in each regression tree, and calculating the mean to generate the average amount of impurity reduction of the meteorological feature. Calculate the average reduction in impurity for all meteorological features. Define the meteorological features with the highest average reduction in impurity as important meteorological features for the feature selector.

[0010] Furthermore, determining the number of meteorological clusters includes: acquiring feature sets of several historical days to generate historical feature sets, and performing clustering by calculating the similarity between several historical feature sets. The number of clusters corresponds to the number of meteorological clusters.

[0011] Furthermore, the labels for the meteorological clusters include clear sky and high temperature type, clear sky and cool type, cloudy, high temperature and high humidity type, and cloudy and rainy, low temperature and high humidity type.

[0012] Furthermore, the construction of the power output prediction model includes: acquiring feature sets and actual photovoltaic power output data for several historical days; dividing the feature sets and corresponding actual photovoltaic power output data based on the similarity between the feature sets and each meteorological family to generate feature data sets for each meteorological family; and constructing power output prediction models for different meteorological families using the feature data sets of different meteorological families.

[0013] Furthermore, the initial predicted value is optimized based on a correction strategy to generate the final predicted value; The correction strategy includes: acquiring air quality forecast data for the forecast day, retrieving the meteorological correction model corresponding to the meteorological cluster to which the forecast day belongs, the meteorological correction model determining the deviation value based on the air quality forecast data for the forecast day, and adding the deviation value to the power output forecast value to generate the final forecast value.

[0014] Furthermore, the air quality forecast data includes PM2.5, PM10, and AQI.

[0015] Furthermore, the construction of the meteorological correction model includes: acquiring the actual photovoltaic output value, predicted photovoltaic output value, air quality data and meteorological clusters for several historical days, and calculating the difference between the actual photovoltaic output value and the predicted photovoltaic output value within the historical days to generate a deviation value; Air quality data and corresponding deviation values ​​are divided into meteorological clusters to generate meteorological data sets for different meteorological clusters. Meteorological correction models for different meteorological clusters are then constructed using these meteorological data sets.

[0016] The beneficial effects of this invention are: By classifying complex weather changes (dividing them into multiple weather clusters), different power output prediction models are corresponding to different types of complex weather changes. When predicting photovoltaic power output, based on the actual weather changes on the prediction day, the power output prediction model corresponding to the weather cluster most similar to the weather changes is used, thus achieving efficient prediction of photovoltaic power output under complex weather conditions.

[0017] By setting a feature set containing several meteorological features, including actual meteorological data of the day before the forecast date, meteorological forecast data of the forecast date, and meteorological forecast data of the days after the forecast date, the photovoltaic output of the forecast date can be determined by combining meteorological data from multiple days. This avoids the limitations of meteorological data from a single date and ensures the accuracy of meteorological output prediction even when the meteorological data on the forecast date is inaccurate.

[0018] By setting a feature selector to extract feature data of important meteorological features to generate a feature set, the accuracy and stability of the power output prediction model can be effectively improved.

[0019] By continuously optimizing important meteorological features within the feature selector, the risk of overfitting in the power output prediction model can be reduced, and the adaptability of the power output prediction model to the constantly changing meteorological environment can be improved.

[0020] By setting up a meteorological correction model that corresponds one-to-one with a meteorological cluster, the corresponding deviation value is determined based on the air quality of the forecast day. The initial forecast value is then adjusted based on the deviation value to generate the final forecast value. This effectively reduces the impact of air quality factors on the nonlinear dynamic influence of photovoltaic power output on the forecast results and improves the accuracy of the forecast.

[0021] By combining the power output prediction model corresponding to the meteorological cluster and the meteorological correction model corresponding to the meteorological cluster, and combining the prediction based on multi-day meteorological data, and then correcting the prediction results based on air quality, the adaptability and robustness of the prediction can be enhanced. The combination of models can also further improve the adaptability of photovoltaic power output prediction under complex meteorological scenarios. Attached Figure Description

[0022] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a schematic diagram of a short-to-medium-term photovoltaic power output prediction method based on multi-day meteorological combinations according to the present invention. Detailed Implementation

[0023] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the specification of this invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed terms.

[0025] This invention provides a method for predicting short-to-medium-term photovoltaic power output based on multi-day meteorological combinations, such as... Figure 1 As shown, the process includes acquiring feature data of several meteorological features within the meteorological data corresponding to the forecast date, extracting feature data of important meteorological features based on a feature selector to form a feature set for the forecast date, calculating the similarity between the feature set and several meteorological clusters respectively, retrieving the power output prediction model corresponding to the meteorological cluster with the highest similarity, and the power output prediction model predicting the initial predicted value representing photovoltaic power output based on the feature set of the forecast date.

[0026] The meteorological data corresponding to the forecast date includes the actual meteorological data of the day before the forecast date, the meteorological forecast data of the forecast date itself, and the meteorological forecast data of the days following the forecast date. By combining meteorological data from multiple days to form a feature set for the forecast date, the limitations of forecasts caused by meteorological data from a single date can be avoided. Even when the meteorological data on the forecast date is inaccurate, the accuracy of subsequent meteorological output forecasts can be improved.

[0027] Meteorological data includes temperature, humidity, wind speed, cloud cover, and solar irradiance. Meteorological data cleaning can improve the validity and accuracy of the data. The cleaning process includes: removing obviously abnormal meteorological feature values ​​from the meteorological data, and then filling in missing meteorological feature values ​​based on the difference in timestamps between adjacent meteorological feature values. One embodiment of this application uses interpolation or mean values ​​to fill in missing meteorological feature values.

[0028] To facilitate the alignment of temperature, humidity, wind speed, cloud cover, and solar irradiance data, the frequency of each data type is uniformly adjusted based on the timestamps of the data, ensuring consistency. One embodiment of this application specifies that the start and end times of the timestamps for temperature, humidity, wind speed, cloud cover, and solar irradiance data are the same, and the difference between adjacent timestamps is 15 minutes.

[0029] Meteorological features include temperature, humidity, wind speed, cloud cover, and solar irradiance, along with their respective mean, variance, maximum, and minimum values. This data can fully capture the daily trends in temperature, humidity, wind speed, cloud cover, and solar irradiance, and comprehensively acquire the features contained in the meteorological data.

[0030] One embodiment of this application is as follows: the calculation window and sliding step size of the features are both set to one day. Therefore, the meteorological features corresponding to the forecast day include the mean, variance, maximum value and minimum value of temperature, humidity, wind speed, cloud cover and solar irradiance on the day before the forecast, the mean, variance, maximum value and minimum value of temperature, humidity, wind speed, cloud cover and solar irradiance on the forecast day, and the mean, variance, maximum value and minimum value of temperature, humidity, wind speed, cloud cover and solar irradiance on the day after the forecast.

[0031] Feature selectors are used to filter out the meteorological features most relevant to photovoltaic power output during the forecast day, which can save computing power for each model while effectively improving the forecast accuracy of the power output prediction model.

[0032] The determination of important meteorological features within the feature selector includes: acquiring feature data and actual photovoltaic power output data of several meteorological features from historical daily meteorological data, and constructing several regression trees using the Random Forest Regression Algorithm (RF-XGBoost). The reduction in impurity caused by a certain meteorological feature at each regression tree's split node is calculated (the greater the reduction in impurity, the greater the impact on the prediction results of the regression tree), and the average reduction in impurity for each regression tree is calculated to generate the average reduction in impurity for that meteorological feature. Using the above method, the average reduction in impurity for all meteorological features is calculated, and the meteorological features with the highest average reduction in impurity are defined as important meteorological features of the feature selector.

[0033] One embodiment of this application is: by continuously updating historical days to continuously optimize the types of important meteorological features in the feature selector, the risk of overfitting of the power output prediction model can be effectively reduced, and the adaptability of the power output prediction model under different scenarios can be improved.

[0034] Determining the number of meteorological clusters involves: acquiring feature data of several meteorological features within meteorological data corresponding to several historical days; extracting feature data of important meteorological features based on a feature selector to construct a feature set for the historical days, and defining it as a historical feature set; determining the similarity between different historical feature sets using a clustering algorithm; and clustering the historical feature sets based on the similarity, with the number of clusters corresponding to the number of meteorological cluster types.

[0035] One embodiment of this application is as follows: When clustering a feature set, if the number of clusters is not limited, meteorological clusters may become overly concentrated or overly dispersed. Over-concentration can affect the accuracy of prediction, while over-dispersion can lead to an excessive number of output prediction models, affecting the overall operating efficiency. To avoid over-concentration or over-dispersion of meteorological clusters during clustering, the elbow rule is used to determine the optimal number of meteorological clusters, which can effectively improve the accuracy of clustering.

[0036] To facilitate the differentiation of different weather clusters, labels are set according to the weather conditions corresponding to the weather clusters. One embodiment of this application is as follows: the labels for weather clusters include clear sky and high temperature type, clear sky and cool type, cloudy, high temperature and high humidity type, and cloudy and rainy, low temperature and high humidity type.

[0037] One embodiment of this application is: each meteorological family corresponds to an XGBoost regression model (power output prediction model). The XGBoost model corresponding to each meteorological family only needs to learn the "meteorology-power output" mapping relationship under a specific meteorological pattern, which reduces the complexity of the model while improving the accuracy of the prediction.

[0038] When generating initial forecast values, the similarity between the feature set of the forecast day and each meteorological cluster is determined based on the clustering algorithm (K-Means algorithm). The XGBoost regression model corresponding to the meteorological cluster with the highest similarity is retrieved. The XGBoost regression model receives the feature set of the forecast day and outputs the initial forecast value representing the photovoltaic power output on the forecast day.

[0039] The construction of the power output prediction model includes: obtaining feature sets and corresponding meteorological power output values ​​for several historical days; dividing the feature sets and corresponding actual photovoltaic power output data based on the similarity between the feature sets and each meteorological family to generate feature data sets for each meteorological family; and constructing power output prediction models for different meteorological families using the feature data sets of different meteorological families.

[0040] The objective function for optimizing the power output prediction model consists of the loss function and the regularization term: in: It is the loss function term (such as mean squared error). The loss function term is used to measure the predicted value. Compared with the true value y i differences γ is the regularization term, which controls the complexity of the model. T is the number of leaf nodes in the tree, w is the score of the leaf nodes, and γ and λ are hyperparameters used to penalize complex models and prevent overfitting.

[0041] Additive Training: ŷ i {(t)} =ŷi {(t-1)} +f t (x i ) In the t-th iteration, add a new tree f. t To fit the residuals of the previous prediction, the prediction results are gradually improved until the model converges or reaches the set number of iterations, thus completing the optimization of the output prediction model.

[0042] It is known that air quality has a nonlinear dynamic impact on photovoltaic (PV) output. Ignoring air quality data when predicting PV output will affect the prediction structure. To improve prediction accuracy, the initial prediction values ​​are optimized using a correction strategy to generate the final prediction values, which can reduce prediction errors caused by air quality factors.

[0043] The correction strategy includes: obtaining air quality forecast data for the forecast date; in one embodiment of this application, the air quality forecast data includes PM2.5, PM10, and AQI. Based on the meteorological cluster for the forecast date, a meteorological correction model corresponding to the meteorological cluster is retrieved. The meteorological correction model determines a deviation value based on the air quality forecast data for the forecast date, and the deviation value is added to the output forecast value to generate the final forecast value.

[0044] The construction of the meteorological correction model includes: acquiring the actual photovoltaic power output, predicted photovoltaic power output, air quality data and meteorological clusters for several historical days, calculating the difference between the actual photovoltaic power output and the predicted photovoltaic power output for the historical days to generate a deviation value; grouping the air quality data and corresponding deviation values ​​for several historical days based on the meteorological clusters to generate datasets for different meteorological clusters, and constructing meteorological correction models for different meteorological clusters using the datasets of different meteorological clusters.

[0045] One embodiment of this application involves using the Random Forest (RF) algorithm to train a meteorological correction model. A K-Means clustering algorithm is employed to cluster air quality data (PM2.5, PM10, AQI, etc.). The meteorological correction model learns the mapping relationship between air quality data (PM2.5, PM10, AQI, etc.) and output deviations (deviation values) under specific meteorological patterns. Based on this mapping relationship, different air quality levels are weighted to derive the deviation values. In this way, the meteorological correction model for each meteorological cluster is highly specialized for a single meteorological scenario, resulting in more accurate predictions.

[0046] The specific implementation process and beneficial effects of this application are as follows: Deployment environment: Python is used as the development language, and the output prediction model and meteorological correction model of each meteorological cluster are built based on libraries such as scikit-learn, XGBoost, pandas, and numpy.

[0047] Offline optimization: Regularly (e.g., monthly) retrain the RF feature selector, XGBoost power prediction model, and RF weather correction model using the latest historical daily data, and update the model parameters.

[0048] Online forecasting: The system acquires meteorological data and air quality forecasts for the next few days from the meteorological bureau at regular intervals every day. Based on the RF feature selector, it determines the feature set for the next few days. The XGBoost power output forecasting model and the RF meteorological correction model automatically generate the photovoltaic power output forecast curve (forecast results) for the next few days based on the feature set and the corresponding air quality forecast data.

[0049] The forecast results can be provided to the power grid dispatching system (EMS) and photovoltaic power plant monitoring system via API or file, so that power grid staff can formulate power generation plans for the next few days and optimize dispatching strategies based on the forecast results.

[0050] This application provides technical support for the full consumption of photovoltaic power, reducing the demand for backup power, and lowering operating costs. It can be widely applied in photovoltaic power plant management, grid dispatch optimization, microgrid system operation, and other fields to improve energy utilization efficiency and economy.

[0051] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Any simple modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention shall still fall within the scope of the technical solution of the present invention.

Claims

1. A method for predicting short-to-medium-term photovoltaic power output based on multi-day meteorological combinations, characterized in that, This includes acquiring feature data of several meteorological features within the meteorological data corresponding to the forecast date, and extracting feature data of important meteorological features based on a feature selector to form a feature set for the forecast date. The similarity between the feature set and several meteorological clusters is calculated respectively. The power output prediction model corresponding to the meteorological cluster with the highest similarity is retrieved. The power output prediction model predicts the initial predicted value representing the photovoltaic power output based on the feature set of the prediction day.

2. The method for predicting short-to-medium-term photovoltaic power output based on multi-day meteorological combinations according to claim 1, characterized in that, The meteorological data corresponding to the forecast date includes the actual meteorological data of the day before the forecast date, the meteorological forecast data of the forecast date, and the meteorological forecast data of the days following the forecast date.

3. The method for predicting short-to-medium-term photovoltaic power output based on multi-day meteorological combinations according to claim 1, characterized in that, The meteorological data includes temperature, humidity, wind speed, cloud cover, and solar irradiance. The meteorological characteristics include temperature, humidity, wind speed, cloud cover, and solar irradiance, with their respective mean, variance, maximum, and minimum values.

4. The method for predicting short-to-medium-term photovoltaic power output based on multi-day meteorological combinations according to claim 1, characterized in that, The determination of important meteorological features in the feature selector includes: acquiring feature data of several meteorological features and actual photovoltaic power output data in meteorological data corresponding to several historical days, and constructing several regression trees based on the random forest regression algorithm; calculating the amount of impurity reduction brought by a certain meteorological feature at its split node in each regression tree, and calculating the mean to generate the average amount of impurity reduction of the meteorological feature. Calculate the average reduction in impurity for all meteorological features. Define the meteorological features with the highest average reduction in impurity as the important meteorological features of the feature selector.

5. The method for predicting short-to-medium-term photovoltaic power output based on multi-day meteorological combinations according to claim 1, characterized in that, The determination of the number of meteorological clusters includes: obtaining feature sets of several historical days to generate historical feature sets, and performing clustering by calculating the similarity between several historical feature sets. The number of clusters corresponds to the number of meteorological clusters.

6. The method for predicting short-to-medium-term photovoltaic power output based on multi-day meteorological combinations according to claim 1, characterized in that, The weather clusters are tagged as follows: clear sky and high temperature, clear sky and cool temperature, cloudy, high temperature and high humidity, and overcast and rainy, low temperature and high humidity.

7. The method for predicting short-to-medium-term photovoltaic power output based on multi-day meteorological combinations according to claim 1, characterized in that, The construction of the power output prediction model includes: acquiring feature sets and actual photovoltaic power output data for several historical days; dividing the feature sets and corresponding actual photovoltaic power output data based on the similarity between the feature sets and each meteorological family to generate feature data sets for each meteorological family; and constructing power output prediction models for different meteorological families using the feature data sets of different meteorological families.

8. The method for predicting short-to-medium-term photovoltaic power output based on multi-day meteorological combinations according to claim 1, characterized in that, The initial predicted value is optimized based on a correction strategy to generate the final predicted value; The correction strategy includes: acquiring air quality forecast data for the forecast day, retrieving the meteorological correction model corresponding to the meteorological cluster to which the forecast day belongs, the meteorological correction model determining the deviation value based on the air quality forecast data for the forecast day, and adding the deviation value to the power output forecast value to generate the final forecast value.

9. A method for predicting short-to-medium-term photovoltaic power output based on multi-day meteorological combinations according to claim 8, characterized in that, The air quality forecast data includes PM2.5, PM10, and AQI.

10. A method for predicting short-to-medium-term photovoltaic power output based on multi-day meteorological combinations according to claim 8, characterized in that, The construction of the meteorological correction model includes: acquiring the actual photovoltaic power output, predicted photovoltaic power output, air quality data and meteorological clusters for several historical days, and calculating the difference between the actual photovoltaic power output and the predicted photovoltaic power output within the historical days to generate a deviation value; Air quality data and corresponding deviation values ​​are divided into meteorological clusters to generate meteorological data sets for different meteorological clusters. Meteorological correction models for different meteorological clusters are then constructed using these meteorological data sets.

Images