A static satellite cloud cover short- and nowcast method and system for photovoltaic power generation

Abstract

This invention discloses a short-term cloud cover prediction method for photovoltaic power generation using geostationary satellites, comprising the following steps: predicting data for the next 4 hours using a neural network; inputting the predicted data into a fast cloud detection algorithm to obtain cloud cover data; correcting the cloud cover data by combining the cloud cover detection results from the operational algorithm with cloud cover observation data from meteorological stations to adjust the parameters of the cloud detection algorithm; and providing a short-term cloud cover prediction system for photovoltaic power generation using geostationary satellites, comprising a neural network algorithm, a fast cloud detection algorithm, and a cloud cover data correction algorithm. The neural network algorithm includes a data acquisition module, a data quality control module, a data normalization processing module, a PreDRNN++ network model, and a predicted data inverse normalization module. The advantages of this invention are: the cloud detection process is rapid and efficient, enabling quick generation of cloud cover data; the prediction system uses the PreDRNN++ network model, considering the changing characteristics of the data in both time and space, resulting in short-term prediction results that are very close to the actual values.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic power generation prediction technology, specifically to a geostationary satellite cloud cover short-term prediction method and system for photovoltaic power generation. Background Technology

[0002] Existing cloud-based detection algorithms are all global detection methods, which require a large amount of auxiliary data such as numerical forecast fields and radiative transfer data. This results in a large computational load and is time-consuming and labor-intensive.

[0003] Currently, there is no cloud detection algorithm that is practical for local areas. However, photovoltaic (PV) stations often only need cloud data in a small area above the PV station. Using business algorithms to calculate this data is time-consuming and costly, and cannot provide timely cloud data for PV stations.

[0004] Existing cloud detection algorithms cannot provide timely and accurate cloud data for photovoltaic (PV) stations. For small areas like PV stations, there is an urgent need for a fast and efficient cloud detection algorithm to provide cloud data. Summary of the Invention

[0005] To address the aforementioned problems, this invention provides a geostationary satellite short-term cloud cover prediction method and system for photovoltaic power generation. The aim is to provide a fast, efficient, and accurate cloud detection algorithm for photovoltaic stations, which are localized and small-scale facilities, and to combine it with a neural network prediction system to predict cloud cover.

[0006] To solve the above-mentioned technical problems, the technical solution provided by this invention is: a geostationary satellite cloud cover short-term prediction method for photovoltaic power generation, comprising the following steps:

[0007] Step 1: The neural network predicts data for the next 4 hours. The detailed steps are as follows:

[0008] S1. Data Acquisition: Acquire satellite data for a period of time prior to the prediction period. This data segment is used as input to the neural network, and its duration is adjusted appropriately based on the duration of the prediction data.

[0009] S2. According to quality control: satellite data with missing phase elements and erroneous satellite data are removed;

[0010] S3. Data normalization processing: Normalize the satellite data from different channels to facilitate processing by the network model;

[0011] S4. Input the data into the predrnn++ network model to obtain the prediction data for the corresponding duration;

[0012] S5. The predicted data is denormalized to give it physical meaning;

[0013] Step 2: Input the predicted data into the fast cloud detection algorithm to obtain cloud cover data. The specific detailed steps are as follows:

[0014] S1. Select the surface type of the photovoltaic station. This algorithm has ordinary surface and desert surface types.

[0015] S2. Input the predicted data for cloud detection. The detection directions include: high cloud detection at infrared brightness temperature of 7.0um, thin cirrus cloud detection at infrared brightness temperatures of 11.2um and 12.3um, low cloud detection at infrared brightness temperatures of 11.2um and 3.9um, visible light reflectance detection at 0.65um, and radiation detection at 0.86 and 0.65 reflectance. After these five detection directions, each phase cell will obtain a final threshold c.

[0016] S3. After the five tests are completed, the final threshold c is obtained. Based on c, the situation of the corresponding phase cell is determined, which is divided into four situations: clear sky, possibly clear sky, possibly cloudy, and cloudy.

[0017] S4. Obtain the overall cloud coverage data within the area above the photovoltaic station;

[0018] Step 3: Cloud cover data correction. Combine the cloud cover detection results of the business algorithm with the cloud cover observation data of the meteorological observation station to correct the cloud detection algorithm parameters. The correction process requires a lot of experiments. In the specific prediction process, the cloud detection algorithm with corrected parameters can be used directly.

[0019] The correction algorithm compares and verifies the detection results of the business algorithm with those of this method. It checks whether there are significant differences in the cloud cover data detected at the same time. If there are significant differences, the parameters corresponding to the cloud detection algorithm will be adjusted. Finally, a set of parameters with the highest detection accuracy will be found for the photovoltaic station. Cloud cover data from meteorological observation stations can also be used for verification.

[0020] A geostationary satellite cloud cover short-term prediction system for photovoltaic power generation includes a neural network algorithm, a fast cloud detection algorithm, and a cloud cover data correction algorithm. The neural network algorithm includes a data acquisition module, a data quality control module, a data normalization processing module, a Predrnn++ network model, and a prediction data inverse normalization module.

[0021] The advantages of this invention compared to existing technologies are: the cloud detection process is fast and efficient, does not require a large amount of auxiliary data, and can quickly obtain cloud volume data; the prediction system adopts the latest PreDRNN++ network model, which considers the changing characteristics of data in terms of time and space, and the short-term prediction effect is very close to the true value. Attached Figure Description

[0022] Figure 1 This is a system diagram of a geostationary satellite cloud cover short-term prediction system for photovoltaic power generation according to the present invention.

[0023] Figure 2 This is a diagram of the predrnn++ network model of a geostationary satellite cloud cover short-term prediction system for photovoltaic power generation according to the present invention.

[0024] Figure 3 This is a schematic diagram of a fast cloud detection algorithm for a geostationary satellite cloud cover short-term prediction system for photovoltaic power generation, as described in this invention. Detailed Implementation

[0025] The present invention will now be described in further detail with reference to the accompanying drawings.

[0026] Example

[0027] A geostationary satellite cloud cover short-term prediction method for photovoltaic power generation includes the following steps:

[0028] Step 1: Use a neural network to predict data for the next 4 hours (the specific duration can be chosen arbitrarily, but prediction accuracy is higher within 4 hours, so 4 hours is preferred). The detailed steps are as follows:

[0029] S1: Data Acquisition: Acquire satellite data for a period of time prior to the prediction period. This data segment is used as input to the neural network, and its duration can be adjusted appropriately based on the duration of the prediction data.

[0030] S2: Data Quality Control: Remove satellite data with missing phase cells and erroneous satellite data, as detailed below:

[0031] (1) Extract the satellite data of the location of the photovoltaic station at 4km per pixel, with the entire range being 32x32 (32 pixels in both length and width);

[0032] (2) Read satellite data: visible light three channels 0.64um, 0.86um, 1.6um; infrared four channels 3.9um, 7.0um, 11.2um, 12.3um;

[0033] (3) During the reading process, it was found that some phase data were missing or did not conform to the actual situation. This part of the data was removed.

[0034] S3: Data Normalization Processing: Satellite data from different channels are normalized separately to facilitate processing by the network model. The normalization formula is as follows:

[0035] X_train (array) is the processed input data; X_train_max is the maximum value in the array; X_train_min is the minimum value in the array; X_train1 is the normalized array:

[0036] X_train1=(X_train-X_train_min) / (X_train_max-X_train_min)

[0037] After normalization, the array data ranges from 0 to 1, which makes it easier for neural networks to extract features from the data;

[0038] S4: Input the data into the Predrnn++ network model to obtain the prediction data for the corresponding duration. The Predrnn++ network model will divide the input data into a training set and a validation set. The training process will find the features of the input data and generate a model. The validation set is used to verify whether the model is suitable for the input data and optimize the model. After the model is generated, it can simulate the data for a future period of time based on the input data to achieve the purpose of prediction.

[0039] S5: Denormalize the predicted data to give it physical meaning:

[0040] After the network model finishes running, the output data is in the same form as the input X_train1, and is not real satellite data. Therefore, it is necessary to perform inverse normalization to make the data have physical meaning like real satellite data.

[0041] Inverse normalization formula: X_train=X_train1*(X_train_max-X_train_min)+X_train_min

[0042] Step 2: Input the predicted data into the fast cloud detection algorithm to obtain cloud cover data, refer to... Figure 3 The specific steps are as follows:

[0043] S1: Select the surface type of the photovoltaic station site. This algorithm has ordinary surface and desert surface types.

[0044] S2: Input the predicted data for cloud detection, with the following 5 aspects:

[0045] (1) Infrared brightness temperature 7.0um high cloud detection;

[0046] (2) Detection of thin cirrus clouds with infrared brightness temperatures of 11.2µm and 12.3µm;

[0047] (3) Infrared brightness temperature of 11.2µm and 3.9µm for low cloud detection;

[0048] (4) Detection of visible light reflectance at 0.65µm;

[0049] (5) Detection of reflectivity at 0.86 and 0.65;

[0050] After these five detection methods, each phase cell will obtain a final threshold c;

[0051] S3: After the five tests are completed, the final threshold c is obtained. Based on c, the situation of the corresponding phase cell is determined, which is divided into four situations: clear sky, possibly clear sky, possibly cloudy, and cloudy.

[0052] When c > 0.99, the sky is considered clear.

[0053] When 0.99 > c > 0.95, it is judged that the sky may be clear.

[0054] When 0.95 > c > 0.66, it is determined that there may be clouds;

[0055] When c < 0.66, it is determined that there is cloud cover.

[0056] S4: Obtain the overall cloud coverage data within the area above the photovoltaic station:

[0057] Cloud coverage rate = (Number of cells with clouds in the photovoltaic station area + Number of cells that may have clouds in the photovoltaic station area) / Total number of cells in the photovoltaic station area

[0058] Step 3: Cloud cover data correction. Combine the cloud cover detection results of the business algorithm with the cloud cover observation data of the meteorological observation station to correct the cloud detection algorithm parameters. The correction process requires a lot of experiments. In the specific prediction process, the cloud detection algorithm with corrected parameters can be used directly.

[0059] The correction algorithm compares and verifies the detection results of the business algorithm with those of this method. It checks whether there are significant differences in the cloud cover data detected at the same time. If there are significant differences, the parameters corresponding to the cloud detection algorithm will be adjusted, and finally a set of parameters with the highest detection accuracy will be used for the photovoltaic station. Cloud cover data from meteorological observation stations can also be used for verification.

[0060] Explanation of terms involved in this invention:

[0061] Fast cloud detection algorithm: It infers the distribution characteristics of clouds based on satellite visible light and infrared channel data.

[0062] Neural networks: By combining low-level features to form more abstract high-level representations of attribute categories or features, they discover distributed feature representations of data.

[0063] Prediction system: Input satellite data, simulate prediction data through neural network, and then input the prediction data into a fast cloud detection algorithm to obtain cloud prediction amount.

[0064] The advantages of this invention are: Timeliness: For short-term photovoltaic forecasting, the timeliness of cloud data is important. This algorithm can provide cloud data for the photovoltaic station for the next few hours in a short period of time; Low computational load: The business algorithm requires huge computing resources to run once, and it is very expensive to provide cloud data for such a small area as a photovoltaic station, which is not economical; High accuracy: The accuracy of cloud detection and prediction is maintained at over 80%.

[0065] This invention and its embodiments have been described, but this description is not restrictive. The accompanying drawings are only one embodiment of the invention, and the actual structure is not limited thereto. In short, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of this invention, such designs should fall within the protection scope of this invention.

Claims

1. A geostationary satellite short-term cloud cover prediction method for photovoltaic power generation, characterized in that, Includes the following steps: Step 1: The neural network predicts data for the next 4 hours. The detailed steps are as follows: S1. Data Acquisition: Acquire satellite data for a period of time prior to the prediction period. This data segment is used as input to the neural network, and its duration is adjusted appropriately based on the duration of the prediction data. S2. According to quality control: satellite data with missing phase elements and erroneous satellite data are removed; S3. Data normalization processing: Normalize the satellite data from different channels to facilitate processing by the network model; S4. Input the data into the predrnn++ network model to obtain the prediction data for the corresponding duration; S5. The predicted data is denormalized to give it physical meaning; Step 2: Input the predicted data into the fast cloud detection algorithm to obtain cloud cover data. The specific detailed steps are as follows: S1. Select the surface type of the photovoltaic station. This algorithm has ordinary surface and desert surface types. S2. Input the predicted data for cloud detection. The detection directions include: high cloud detection at infrared brightness temperature of 7.0um, thin cirrus cloud detection at infrared brightness temperatures of 11.2um and 12.3um, low cloud detection at infrared brightness temperatures of 11.2um and 3.9um, visible light reflectance detection at 0.65um, and radiation detection at 0.86 and 0.65 reflectance. After these five detection directions, each phase cell will obtain a final threshold c. S3. After the five tests are completed, the final threshold c is obtained. Based on c, the situation of the corresponding phase cell is determined, which is divided into four situations: clear sky, possibly clear sky, possibly cloudy, and cloudy. S4. Obtain the overall cloud coverage data within the area above the photovoltaic station; Step 3: Cloud cover data correction. Combine the cloud cover detection results of the business algorithm with the cloud cover observation data of the meteorological observation station to correct the cloud detection algorithm parameters. The correction process requires a lot of experiments. In the specific prediction process, the cloud detection algorithm with corrected parameters can be used directly. The correction algorithm compares and verifies the detection results of the business algorithm with those of this method. It checks whether there are significant differences in the cloud cover data detected at the same time. If there are significant differences, the parameters corresponding to the cloud detection algorithm will be adjusted. Finally, a set of parameters with the highest detection accuracy will be found for the photovoltaic station. Cloud cover data from meteorological observation stations can also be used for verification.

2. A geostationary satellite cloud cover short-term prediction system for photovoltaic power generation, executing the geostationary satellite cloud cover short-term prediction method for photovoltaic power generation as described in claim 1, characterized in that: It includes a data acquisition module, a data quality control module, a data normalization processing module, a Predrnn++ network model, and a prediction data inverse normalization module.

Images