Regional distributed photovoltaic prediction method and device based on space-time convolutional neural network

Abstract

The invention provides a regional distributed photovoltaic prediction method and device based on a space-time convolutional neural network. The method comprises the following steps: S1, collecting actually measured meteorological data and actual output data of each distributed photovoltaic station at a fixed time interval; s2, correcting and cleaning the predictive weather; s3, carrying out regional distributed photovoltaic clustering; s4, performing model construction and training to obtain a respective photovoltaic output prediction model of each category; and S5, analyzing the prediction result of the respective photovoltaic output prediction model of each category by using a weighted linear regression algorithm to obtain the weight of each prediction result. According to the invention, based on the spatio-temporal data of various stations of regional distributed photovoltaic, the spatio-temporal convolutional neural network model is adopted to carry out short-term output prediction on various stations, so that the whole regional distributed photovoltaic can be predicted more accurately.

Description

Technical Field

[0001] The present invention belongs to the field of new energy power technology, and in particular relates to a method and device for predicting power of regional distributed photovoltaics based on spatiotemporal convolutional neural networks. Background Art

[0002] As the new power system with new energy as the main body has become an industry consensus, distributed photovoltaics, as an indispensable part of the emerging power system, are particularly remarkable for their characteristics of nearby power generation, nearby grid connection, nearby conversion, and nearby use, which effectively solves the problem of power loss during voltage boosting and long-distance transportation, thereby significantly improving energy utilization efficiency. Therefore, it is crucial to improve the accuracy of short-term output forecasts for distributed photovoltaics.

[0003] At present, the installed capacity of distributed photovoltaic power generation is gradually increasing. In some areas, its installed capacity is already on par with that of centralized photovoltaic power generation. The accuracy of its short-term prediction has had a significant impact on the dispatch and absorption of the power grid.

[0004] Due to the dispersed layout of distributed photovoltaics, the environmental conditions, equipment status and operating data of each area may be different, making it difficult to predict its short-term output. Therefore, distributed photovoltaic predictions often use the model station conversion method for prediction modeling, which means that distributed photovoltaics can only predict output based on the model station as the benchmark meteorology. However, the power generation capacity of the photovoltaic system is significantly affected by a variety of meteorological factors, including cloud movement and thickness (high, medium and low clouds), radiation intensity, humidity, temperature, etc. Traditional model building methods usually only analyze the impact of photovoltaic power generation on the meteorological conditions of a specific site. Since the model stations are usually far apart, this method is difficult to accurately reflect the impact of cloud movement on the distributed photovoltaic system, which in turn affects the accuracy of photovoltaic power generation prediction in the entire region. Summary of the invention

[0005] The present invention proposes a regional distributed photovoltaic prediction method and device based on spatiotemporal convolutional neural network, which can make short-term output predictions for various sites based on the spatiotemporal data of various sites by using a spatiotemporal convolutional neural network model, thereby making a more accurate prediction of the entire regional distributed photovoltaic.

[0006] To achieve the above object, the technical solution of the present invention is achieved as follows:

[0007] A regional distributed photovoltaic prediction method based on spatiotemporal convolutional neural network, comprising:

[0008] S1. Collect measured meteorological data and actual output data of each distributed photovoltaic site at fixed time intervals;

[0009] S2. Correction and cleaning of forecast meteorological data: The forecast meteorological data of each distributed photovoltaic station is corrected by using the measured meteorological data, and the meteorological features with large deviations in the forecast meteorological data are eliminated by the root mean square error (RMSE) evaluation method to obtain the corrected forecast meteorological data;

[0010] S3. Regional distributed photovoltaic clustering: Four types of clustering are performed, including (1) clustering based on the cleaned forecast meteorological data of each site using the K-means++ clustering method; (2) clustering based on the installed capacity of distributed sites using the hierarchical clustering method; (3) clustering based on the coordinates of distributed sites using the density clustering method; (4) clustering based on the measured output data of each site using the hierarchical clustering method; and the categories of the four clusters are obtained;

[0011] S4, model construction and training: Based on the corrected forecast meteorological data and the actual output data of each distributed photovoltaic site, combined with the respective categories of the four clusters in step S3, a spatiotemporal convolutional neural network algorithm is used to establish and train a photovoltaic output prediction model in each category to obtain a photovoltaic output prediction model for each category;

[0012] S5. Analyze the prediction results of the photovoltaic output prediction model of each category using a weighted linear regression algorithm to obtain the weight of each prediction result.

[0013] Furthermore, in step S1, measured meteorological data are collected at fixed time intervals from distributed monitoring points of each distributed photovoltaic site using detection equipment, and the measured meteorological data include cloud thickness, solar radiation, humidity, and temperature. The detection equipment includes a laser cloud height meter, a solar irradiance meter, and temperature and humidity sensors.

[0014] Further, in step S2, the correction method includes:

[0015] S201, calculating the error between the predicted value of the predicted meteorological data and the measured value of the measured meteorological data, performing regression analysis using the linear relationship between the data value of the measured meteorological data and the data value of the predicted meteorological data, and establishing a correction model;

[0016] S202, calculating the root mean square difference RMSE between the predicted value and the measured value for each meteorological data, setting a threshold by analyzing the distribution of RMSE, and eliminating the predicted meteorological data exceeding the threshold.

[0017] Furthermore, in step S4, the establishment and training of the photovoltaic output prediction model specifically includes:

[0018] S401. Define a spatiotemporal convolutional neural network, including spatial convolution, temporal convolution, and fusion layers, use a mean square error loss function, and use Adam to optimize hyperparameters;

[0019] S402, the corrected forecast meteorological data is standardized and divided into time series segments, and then input into a spatiotemporal convolutional neural network for training, wherein the spatial convolution and the temporal convolution process the features in the spatial and temporal dimensions respectively, and the processed features are input into a fusion layer.

[0020] Furthermore, the weighted linear regression algorithm in step S5 includes:

[0021] A weighted linear regression model is defined, and the prediction results output by the photovoltaic output prediction model of each category are input into the weighted linear regression model as features. The model is evaluated by MSE, and finally the weights of the prediction results output by the photovoltaic output prediction model of each category are output.

[0022] On the other hand, the present invention also proposes a regional distributed photovoltaic prediction device based on spatiotemporal convolutional neural network, comprising:

[0023] Data collection module: collects measured meteorological data and actual output data of each distributed photovoltaic site at fixed time intervals;

[0024] Data correction module: used for forecast meteorological correction and cleaning, using the measured meteorological data to correct the forecast meteorological data of each distributed photovoltaic site, and eliminating the meteorological features with large deviations in the forecast meteorological data through the RMSE evaluation method to obtain the corrected forecast meteorological data;

[0025] Clustering module: used for regional distributed photovoltaic clustering: four types of clustering are performed, including (1) clustering based on the cleaned forecast meteorological data of each site using the K-means++ clustering method; (2) clustering based on the installed capacity of distributed sites using the hierarchical clustering method; (3) clustering based on the coordinates of distributed sites using the density clustering method; (4) clustering based on the measured output data of each site using the hierarchical clustering method; and obtaining the categories of the four clusters;

[0026] Model building and training module: Based on the revised forecast meteorological data and the actual output data of each distributed photovoltaic site, combined with the respective categories of the four clusters in the clustering module, a spatiotemporal convolutional neural network algorithm is used in each category to establish and train a photovoltaic output forecast model, and a photovoltaic output forecast model for each category is obtained;

[0027] Analysis module: The prediction results of the photovoltaic output prediction model of each category are analyzed using a weighted linear regression algorithm to obtain the weight of each prediction result.

[0028] Furthermore, in the data acquisition module, detection equipment is used to collect measured meteorological data at fixed time intervals from distributed monitoring points of each distributed photovoltaic site. The measured meteorological data include cloud thickness, solar radiation, humidity, and temperature. The detection equipment includes a laser cloud height meter, a solar irradiance meter, and temperature and humidity sensors.

[0029] Furthermore, the data correction module includes:

[0030] Correction model unit: calculate the error between the predicted value of the predicted meteorological data and the measured value of the measured meteorological data, perform regression analysis using the linear relationship between the data value of the measured meteorological data and the data value of the predicted meteorological data, and establish a correction model;

[0031] Analysis unit: Calculate the root mean square difference (RMSE) between the predicted value and the measured value for each meteorological data, set a threshold by analyzing the distribution of RMSE, and eliminate the predicted meteorological data that exceeds this threshold.

[0032] Furthermore, the model building and training modules include:

[0033] Definition unit: Define the spatiotemporal convolutional neural network, including spatial convolution, temporal convolution, and fusion layers, use the mean square error loss function, and use Adam to optimize hyperparameters;

[0034] Training unit: The corrected forecast meteorological data is standardized and divided into time series segments, and then input into the spatiotemporal convolutional neural network for training, wherein the spatial convolution and the temporal convolution process the features in the spatial and temporal dimensions respectively, and the processed features are input into the fusion layer.

[0035] Furthermore, the analysis module includes:

[0036] A weighted linear regression model is defined, and the prediction results output by the photovoltaic output prediction model of each category are input into the weighted linear regression model as features. The model is evaluated by MSE, and finally the weights of the prediction results output by the photovoltaic output prediction model of each category are output.

[0037] Compared with the prior art, the present invention has the following beneficial effects:

[0038] 1. The present invention adopts the RMSE evaluation method to eliminate meteorological features with large prediction deviations, thereby reducing the deviation of distributed photovoltaic prediction output caused by meteorological deviations.

[0039] 2. The present invention clusters regional distributed photovoltaics in different ways, adopts a spatiotemporal convolutional neural network algorithm to predict power for each clustering method, and performs weighted prediction on each clustering result through weighted linear regression to obtain the final result, thereby improving the generalization ability of the model and reducing the regional power prediction deviation caused by meteorological forecast deviation.

[0040] 3. The present invention adopts spatiotemporal convolutional neural network to extract spatiotemporal features. It has the characteristics of powerful spatiotemporal feature modeling ability, automatic feature learning ability, high prediction accuracy and robustness, etc. It can accurately capture the movement and changes of clouds and the influence of other meteorological factors, and provide higher accuracy and reliability for the power generation prediction of regional distributed photovoltaic systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 is a schematic diagram of a flow chart of an embodiment of the present invention;

[0042] Figure 2 4 is a comparison chart of prediction results of an embodiment of the present invention. DETAILED DESCRIPTION

[0043] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments may be combined with each other.

[0044] The main design idea of ​​the present invention is to cluster regional distributed photovoltaics from multiple angles, then obtain the spatiotemporal data of various types of sites, and use the spatiotemporal convolutional neural network model to predict the output of each type of site, thereby making a more accurate prediction of the entire regional distributed photovoltaics.

[0045] In order to make the purpose and features of the present invention more obvious and understandable, the specific implementation methods of the present invention are further described below in conjunction with the accompanying drawings.

[0046] like Figure 1 As shown, the method proposed in this embodiment specifically includes:

[0047] S1. Data collection: From multiple distributed monitoring points at each site of regional distributed photovoltaics, laser cloud height meters, solar irradiance meters, temperature, humidity and other sensors are used to collect measured meteorological data (including cloud thickness, solar radiation, humidity, temperature, etc.) and distributed photovoltaic actual output data at fixed time intervals.

[0048] S2. Correction and cleaning of forecasted meteorological data: Obtain forecasted meteorological data of the same period as the measured meteorological data obtained in step S1, correct the forecasted meteorological data based on the measured meteorological data, and eliminate meteorological features with large deviations through the RMSE evaluation method to reduce the forecast output deviation caused by the deviation of the forecasted meteorological data.

[0049] The specific method of this step is:

[0050] 1. Calculate the error between the predicted value and the measured value of meteorological data, use the linear relationship between the measured value and the predicted value for regression analysis, and establish a correction model;

[0051] 2. Calculate the root mean square error (RMSE) between the predicted value and the measured value of the meteorological data, set a threshold by analyzing the distribution of RMSE, and remove the predicted meteorological data that exceeds this threshold; finally, obtain the corrected predicted meteorological data;

[0052] 3. Use the revised forecast meteorological data for subsequent analysis and model prediction to improve the accuracy of the forecast.

[0053] The above method generates input data suitable for training the spatiotemporal convolutional neural network (ST-CNN) model through operations such as data cleaning, normalization, and reorganization of spatiotemporal data.

[0054] S3, regional distributed photovoltaic clustering: In this step, four clustering methods are adopted, and four types of clustering are performed respectively, as described below:

[0055] 1. Based on the cleaned forecast meteorological data of each station, the K-means++ clustering method is used to cluster them, specifically:

[0056] 1.1 Standardize weather forecasts;

[0057] 1.2 Use the elbow method to find the number of cluster centers K;

[0058] 1.3 Apply K-means++ for clustering.

[0059] 2. Based on the installed capacity of distributed sites, a hierarchical clustering method is used to cluster them, specifically:

[0060] 2.1 Collect installed capacity data of each site;

[0061] 2.2 Manhattan distance is used to calculate the similarity of installed capacity of each site;

[0062] 2.3 Select average linkage for clustering.

[0063] 3. Based on the distributed site coordinates, the density clustering method is used to cluster them, specifically:

[0064] 3.1 Collect the latitude and longitude information of each site;

[0065] 3.2 Use Euclidean distance to measure the direct distance between sites;

[0066] 3.3 Use K-distance graph to select the appropriate eps value;

[0067] 3.4 Perform density clustering.

[0068] 4. Based on the measured output data of each station, a hierarchical clustering method is used to cluster them, specifically:

[0069] 4.1 Clean the missing values ​​and outliers in the measured output data and standardize the data;

[0070] 4.2 Calculate the distance matrix between each site;

[0071] 4.3 Find similar sites and cluster them.

[0072] S4, model construction and training: Based on the revised forecast meteorological data of each regional distributed photovoltaic site obtained in step S2, the actual output data of each regional distributed photovoltaic site obtained in step S1, and the clustering results (categories) obtained by four clustering methods in step S3, a spatiotemporal convolutional neural network algorithm is used in each category to establish a photovoltaic output prediction model, specifically:

[0073] 1. Define the spatiotemporal convolutional neural network, which mainly includes spatial convolution, temporal convolution, and fusion layers, adopt the mean square error loss function, and use Adam to optimize hyperparameters;

[0074] 2. The weather forecast data is standardized and divided into time series segments, and then input into the spatiotemporal convolutional neural network, where spatial convolution and temporal convolution process the features in the spatial and temporal dimensions respectively. The processed features are input into the fusion layer to obtain the output results.

[0075] Compared with traditional photovoltaic output prediction algorithms and models, the photovoltaic output prediction model in this embodiment includes a temporal convolution layer and a spatial convolution layer, which can simultaneously capture the changes and correlations of meteorological data in time and space. Through convolution operations, the spatiotemporal convolutional neural network ST-CNN can automatically extract the complex relationship between meteorological factors and photovoltaic power output, and reduce the computational complexity by reducing the feature size through pooling operations, while retaining important features, thereby improving the training speed and prediction accuracy of the network.

[0076] S5: Finally, a weighted linear regression algorithm is used for the prediction results output by the photovoltaic output prediction model of each category. Specifically, a weighted linear regression model is defined, the results output by each clustering algorithm are used as feature input models, the quality of the model is evaluated by MSE, and finally the prediction results and the weights of each clustering result are output, so as to obtain the final regional distributed photovoltaic power prediction results.

[0077] In this embodiment, the above prediction method is used to compare the power prediction of the new and old algorithms for a certain distributed project. The results are as follows: Figure 2As shown, the prediction result of the new algorithm proposed in the present invention is the red curve in the figure. Compared with the blue curve of the actual photovoltaic output, the prediction accuracy is 95.2%, while the prediction result of the conventional old algorithm is the green curve in the figure. Compared with the blue curve of the actual photovoltaic output, the prediction accuracy is 88.44%, and the accuracy is improved by 6.76%.

[0078] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A regional distributed photovoltaic prediction method based on spatiotemporal convolutional neural network, characterized in that: include: S1. Collect measured meteorological data and actual output data of each distributed photovoltaic site at fixed time intervals; S2. Correction and cleaning of forecast meteorological data: The forecast meteorological data of each distributed photovoltaic station is corrected by using the measured meteorological data, and the meteorological features with large deviations in the forecast meteorological data are eliminated by the root mean square error (RMSE) evaluation method to obtain the corrected forecast meteorological data; S3. Regional distributed photovoltaic clustering: Four types of clustering are performed, including (1) clustering based on the cleaned forecast meteorological data of each site using the K-means++ clustering method; (2) clustering based on the installed capacity of distributed sites using the hierarchical clustering method; (3) clustering based on the coordinates of distributed sites using the density clustering method; (4) clustering based on the measured output data of each site using the hierarchical clustering method; and the categories of the four clusters are obtained; S4, model construction and training: Based on the corrected forecast meteorological data and the actual output data of each distributed photovoltaic site, combined with the respective categories of the four clusters in step S3, a spatiotemporal convolutional neural network algorithm is used to establish and train a photovoltaic output prediction model in each category to obtain a photovoltaic output prediction model for each category; S5. Analyze the prediction results of the photovoltaic output prediction model of each category using a weighted linear regression algorithm to obtain the weight of each prediction result.

2. The regional distributed photovoltaic prediction method based on spatiotemporal convolutional neural network according to claim 1 is characterized in that: In step S1, measured meteorological data are collected at fixed time intervals from distributed monitoring points of each distributed photovoltaic site using detection equipment, the measured meteorological data including cloud thickness, solar radiation, humidity, and temperature, and the detection equipment includes a laser cloud height meter, a solar irradiance meter, and a temperature and humidity sensor.

3. The regional distributed photovoltaic prediction method based on spatiotemporal convolutional neural network according to claim 1 is characterized in that: In step S2, the correction method includes: S201, calculating the error between the predicted value of the predicted meteorological data and the measured value of the measured meteorological data, performing regression analysis using the linear relationship between the data value of the measured meteorological data and the data value of the predicted meteorological data, and establishing a correction model; S202, calculating the root mean square difference RMSE between the predicted value and the measured value for each meteorological data, setting a threshold by analyzing the distribution of RMSE, and eliminating the predicted meteorological data exceeding the threshold.

4. The regional distributed photovoltaic prediction method based on spatiotemporal convolutional neural network according to claim 1 is characterized in that: In step S4, the establishment and training of the photovoltaic output prediction model specifically includes: S401. Define a spatiotemporal convolutional neural network, including spatial convolution, temporal convolution, and fusion layers, use a mean square error loss function, and use Adam to optimize hyperparameters; S402, the corrected forecast meteorological data is standardized and divided into time series segments, and then input into a spatiotemporal convolutional neural network for training, wherein the spatial convolution and the temporal convolution process the features in the spatial and temporal dimensions respectively, and the processed features are input into a fusion layer.

5. The regional distributed photovoltaic prediction method based on spatiotemporal convolutional neural network according to claim 1 is characterized in that: The weighted linear regression algorithm in step S5 includes: A weighted linear regression model is defined, and the prediction results output by the photovoltaic output prediction model of each category are input into the weighted linear regression model as features. The model is evaluated by MSE, and finally the weights of the prediction results output by the photovoltaic output prediction model of each category are output.

6. A regional distributed photovoltaic prediction device based on spatiotemporal convolutional neural network, characterized in that: include: Data collection module: collects measured meteorological data and actual output data of each distributed photovoltaic site at fixed time intervals; Data correction module: used for forecast meteorological correction and cleaning, using the measured meteorological data to correct the forecast meteorological data of each distributed photovoltaic site, and eliminating the meteorological features with large deviations in the forecast meteorological data through the RMSE evaluation method to obtain the corrected forecast meteorological data; Clustering module: used for regional distributed photovoltaic clustering: four types of clustering are performed, including (1) clustering based on the cleaned forecast meteorological data of each site using the K-means++ clustering method; (2) clustering based on the installed capacity of distributed sites using the hierarchical clustering method; (3) clustering based on the coordinates of distributed sites using the density clustering method; (4) clustering based on the measured output data of each site using the hierarchical clustering method; and obtaining the categories of the four clusters; Model building and training module: Based on the revised forecast meteorological data and the actual output data of each distributed photovoltaic site, combined with the respective categories of the four clusters in the clustering module, a spatiotemporal convolutional neural network algorithm is used in each category to establish and train a photovoltaic output forecast model, and a photovoltaic output forecast model for each category is obtained; Analysis module: The prediction results of the photovoltaic output prediction model of each category are analyzed using a weighted linear regression algorithm to obtain the weight of each prediction result.

7. The regional distributed photovoltaic prediction device based on spatiotemporal convolutional neural network according to claim 6, characterized in that: In the data acquisition module, detection equipment is used to collect measured meteorological data at fixed time intervals from distributed monitoring points of each distributed photovoltaic site. The measured meteorological data includes cloud thickness, solar radiation, humidity, and temperature. The detection equipment includes a laser cloud height meter, a solar irradiance meter, and temperature and humidity sensors.

8. The regional distributed photovoltaic prediction device based on spatiotemporal convolutional neural network according to claim 6, characterized in that: The data correction module includes: Correction model unit: calculate the error between the predicted value of the predicted meteorological data and the measured value of the measured meteorological data, perform regression analysis using the linear relationship between the data value of the measured meteorological data and the data value of the predicted meteorological data, and establish a correction model; Analysis unit: Calculate the root mean square difference (RMSE) between the predicted value and the measured value for each meteorological data, set a threshold by analyzing the distribution of RMSE, and eliminate the predicted meteorological data that exceeds this threshold.

9. The regional distributed photovoltaic prediction device based on spatiotemporal convolutional neural network according to claim 6, characterized in that: Model building and training modules include: Definition unit: Define the spatiotemporal convolutional neural network, including spatial convolution, temporal convolution, and fusion layers, use the mean square error loss function, and use Adam to optimize hyperparameters; Training unit: The corrected forecast meteorological data is standardized and divided into time series segments, and then input into the spatiotemporal convolutional neural network for training, wherein the spatial convolution and the temporal convolution process the features in the spatial and temporal dimensions respectively, and the processed features are input into the fusion layer.

10. The regional distributed photovoltaic prediction device based on spatiotemporal convolutional neural network according to claim 6, characterized in that: The analysis modules include: A weighted linear regression model is defined, and the prediction results output by the photovoltaic output prediction model of each category are input into the weighted linear regression model as features. The model is evaluated by MSE, and finally the weights of the prediction results output by the photovoltaic output prediction model of each category are output.

Images