New energy power station ultra-short-term irradiance prediction method considering space-time dependence

Abstract

The invention discloses a new energy power station ultra-short-term irradiance prediction method considering space-time dependence. The method comprises the following specific steps: S1, preprocessing irradiance actual measurement historical data and earth surface irradiance data of a surrounding area of a new energy power station inversed by satellite multi-channel data; s2, constructing a unified prediction model based on a space-time diagram convolutional neural network; the step S2 comprises: S2-1, constructing graph structure data taking a target power station and a space-time correlation power station as nodes, and a prediction model of a space-time graph convolutional neural network taking the graph structure data as input and the irradiance of the target power station as output; s2-2, training by using a prediction model of a space-time diagram convolutional neural network to obtain a unified prediction model; s3, performing ultra-short-term irradiance prediction by using the obtained unified prediction model; the unified prediction model constructed by the invention is high in prediction precision, and can effectively overcome the problem that a traditional irradiance prediction model cannot quickly fit instantaneous change for a sudden cloud layer shielding event in different weather states.

Description

Technical Field

[0001] The present invention belongs to the field of new energy technology, and relates to a method for predicting ultra-short-term irradiance of a new energy power station, and in particular to a method for predicting ultra-short-term irradiance of a new energy power station taking time and space dependence into consideration. Background Art

[0002] New energy power forecast is to predict the wind power or photovoltaic power sequence of a certain site or region in the future. Its time scale is divided into ultra-short-term, short-term and medium-long term. Ultra-short-term generally refers to 0-4 hours, predicting the wind power or photovoltaic power sequence between 8-12 o'clock; short-term is mainly for intraday or 1-7 days forecast; medium-long term is mainly for quarterly or annual forecast. Accurate forecast is crucial to the balance of supply and demand and economic dispatch of power grid. Among them, ultra-short-term and short-term forecasts help to reasonably reserve spare capacity, adjust unit combination schemes, and optimize power generation plans.

[0003] At present, the research on ultra-short-term power prediction of renewable energy power stations is mainly divided into two steps, namely surface irradiance prediction and surface irradiance-renewable energy power mapping. Since the current surface irradiance-renewable energy power mapping has a high accuracy, but the irradiance prediction still has a large error, especially the tracking and prediction ability of rapidly fluctuating irradiance under non-clear weather conditions needs to be further improved.

[0004] In terms of irradiance prediction models, most ultra-short-term irradiance prediction methods do not distinguish between weather types during the prediction period and use a unified model for prediction. When the predicted period is cloudy or other weather conditions, sudden cloud cover events will cause a sudden change in irradiance, making it impossible to guarantee the accuracy of the prediction. Classification modeling based on different weather types is an effective way to solve this problem, which requires classification and identification of different weather types and prediction of weather types for the next 0-4 hours.

[0005] In addition, in terms of data types, a large number of studies have shown that solar irradiance data has temporal and spatial correlations, while research on ultra-short-term irradiance predictions for new energy power stations is mainly based on numerical weather forecasts, satellite / ground-based cloud maps, and site monitoring data. These data, which are limited to a single spatial location, can only mine the time-dependent information of the data, and it is difficult to obtain the spatial-dependent information of the surrounding areas.

[0006] Therefore, it is also necessary to establish a more accurate ultra-short-term irradiance prediction model for new energy power stations to further improve the accuracy of irradiance prediction. Summary of the invention

[0007] In order to build a more accurate irradiance prediction model for new energy power stations and improve the accuracy of irradiance prediction, the present invention classifies and models different weather types to provide accurate weather pattern information for the prediction period, and at the same time corrects the measured irradiance data, and establishes a unified prediction model based on spatiotemporal graph convolutional neural network.

[0008] The technical solution adopted by the present invention is a method for predicting ultra-short-term irradiance of a new energy power station taking into account time and space dependence. The key lies in the following steps:

[0009] S1. Preprocessing of surface irradiance data around new energy power stations from measured historical data and satellite multi-channel data inversion;

[0010] S2. Construct a unified prediction model based on spatiotemporal graph convolutional neural network;

[0011] The S2 step includes:

[0012] S2-1. Construct a prediction model of a spatiotemporal graph convolutional neural network with graph structure data taking the target power station and spatiotemporally related power stations as nodes, graph structure data as input, and target power station irradiance as output;

[0013] S2-2, using the data preprocessed in step S1 to train the prediction model of the above-mentioned spatiotemporal graph convolutional neural network to obtain a unified prediction model;

[0014] S3. Use the obtained unified prediction model to perform ultra-short-term irradiance prediction.

[0015] Specifically, the spatiotemporal graph convolutional neural network in step S2 is composed of two spatiotemporal convolution blocks and a fully connected layer. The spatiotemporal convolution blocks are used to extract the spatiotemporal correlation of the data. Each spatiotemporal convolution block includes a first time-gated convolution layer at the upper layer, a second time-gated convolution layer at the lower layer, and a spatial graph convolution layer in the middle; the fully connected output layer is used for the final result prediction, specifically, the time dimension of the original output data is merged by convolution of the time dimension, and the final prediction data is output after the merging through a convolution. The prediction data is a graph of the next time dimension.

[0016] More specifically, each spatiotemporal convolution block uses a 1D causal convolution implemented using a gating mechanism in the time dimension. The spatiotemporal convolution block input data p l , output data p according to formula 1 l+1 , Formula 1 is:

[0017]

[0018] In formula 1, and denote the convolution kernels on the first and second temporal convolutions, Θ l Represents the convolution kernel on the graph convolution, and ReLU represents the rectified linear unit function.

[0019] Furthermore, the above preprocessing includes:

[0020] S1-1. Preprocessing of historical irradiance measurement data includes filling in missing data and identifying, deleting and filling in abnormal data:

[0021] The missing data in the measured historical data are filled in by interpolation technology based on the clear sky index;

[0022] Abnormal data in measured historical data are identified and deleted based on the 3-sigma criterion, and the KNN method is used to fill in missing values;

[0023] S1-2. The preprocessing of the surface irradiance data of the area around the new energy power station inverted from the satellite multi-channel data is corrected by using the irradiance deviation correction model.

[0024] Preferably, the interpolation technology of the clear sky index is to use the irradiance data inverted by satellite to calculate the clear sky index within the timestamp of the missing data, and then fill the gap on the missing timestamp by multiplying the clear sky index in the timestamp with the expected irradiance in clear weather.

[0025] Furthermore, the above correction using the irradiance deviation correction model includes:

[0026] S1-2-1. Establish an irradiance deviation correction model and calculate the correction factor for overlapping data;

[0027] S1-2-2. Use the irradiance deviation correction model to correct each inverted irradiance time series data to obtain corrected inverted irradiance time series data.

[0028] Specifically, step S1-2-1 uses the pre-processed measured irradiance data and the surface irradiance data of the area around the new energy power station inverted by the satellite multi-channel data to correct the difference between the fitting straight line and the straight line, and calculate the correction factors a and b;

[0029] Step S1-2-2 calculates the new estimated value through equation 2, which is:

[0030] y new =y satellite -[(a-1)x ground +b] Formula 2,

[0031] In formula 2, y new Represents the corrected irradiance data, y satelliterepresents the surface irradiance data of the area around the new energy power station inverted by satellite multi-channel data, x ground represents the measured irradiance data after preprocessing, and a and b represent correction factors.

[0032] The beneficial effects of the present invention are:

[0033] This study proposes a new energy power station irradiance prediction model by integrating deep learning methods and spatiotemporal dependency principles.

[0034] In order to avoid missing or abnormal data points caused by interference from factors such as artificial abandonment of light, communication failure, or unit disconnection in actual operation, the measured irradiance data was filled with missing data. The measured historical data of surface irradiance is an important basis for modeling and predictive analysis. However, sometimes due to communication system failures, large sections of data are missing in the collected data, which affects the integrity of the data set. Many studies use linear interpolation and Kalman filters, but these methods are not suitable for data gaps that last for more than one day. For surface irradiance data, if only a small amount of data is missing throughout the day, linear interpolation is used to fill it. It assumes that there is a strong trend in the changes between time series, and the trend changes of the data are fitted to fill it. If there is one or more days of missing data, it is more scientific to use interpolation technology based on the clear sky index. This is because the clear sky index reflects the proportion of solar radiation that is reduced after entering the atmosphere, and this proportion is directly related to the cloudy and sunny state of the weather.

[0035] In addition, in actual operation, due to interference from factors such as artificial abandonment of light, communication failures, or unit disconnection from the grid, there are often abnormal data points in the measured historical data of surface irradiance, and these data cannot reflect the actual operating characteristics of new energy power stations. Common methods for identifying abnormal data include clustering algorithms, Markov chain methods, principal component analysis methods, etc. These algorithms have been applied in the fields of wind farm abandonment abnormal data identification and abnormal load data identification. However, new energy power data has strong periodicity and random volatility, and the above algorithms are not suitable for the identification of new energy abnormal data. This study is more effective in identifying abnormal data based on the 3-sigma criterion.

[0036] The simplest processing method is to directly delete the identified abnormal data, but the reduction in data volume caused by direct deletion may affect the integrity of the new energy output time series and the effect of model training, so abnormal data repair is required. The present invention uses the KNN method to fill in the vacancy value, which is better than the MICE method and the MF method.

[0037] In addition, in data preprocessing, the present invention performs deviation correction on the surface irradiance data of the area around the new energy power station inverted by satellite multi-channel data, thereby further improving the accuracy of the data.

[0038] The present invention realizes the identification and reconstruction of abnormal data of new energy power based on irradiance data, which is conducive to building a more accurate unified prediction model based on spatiotemporal graph convolutional neural network.

[0039] As an important branch of machine learning, deep learning can improve the ability to extract data features by building a deeper and more complex neural network architecture. Compared with traditional shallow neural networks, deep learning can extract deep features contained in data through a mesh hierarchical structure through a series of nonlinear mathematical transformations, thereby performing classification or prediction. The modules commonly used in deep learning are convolutional neural networks, represented by CNN, and temporal neural networks, represented by RNN. CNN is good at extracting spatial information features of data, while RNN is good at extracting temporal feature information of data. In the prior art, some studies have proposed a method of splicing CNN and GRU to build a composite network for research. This network is called spatiotemporal neural network in the field of machine learning, represented by STNN. The spatiotemporal neural network first uses CNN to extract the spatial features of the data, and then uses long short-term memory units or recurrent gating units to extract the temporal dimension features of the data. It has achieved good results in video behavior detection and classification tasks in the field of artificial intelligence. However, the recurrent gating unit is a unidirectional GRU, which can only extract unidirectional features. After CNN extracts features and passes them to RNN, its features should have multi-dimensional attributes. If only one-way GRU is used for regression, some useful information will be lost.

[0040] The present invention combines a graph convolutional neural network model capable of extracting spatial features and a temporal convolutional network capable of extracting temporal features to form a spatiotemporal graph convolutional network framework capable of extracting spatiotemporal features at the same time. The network consists of two spatiotemporal convolutional blocks and a fully connected layer. The spatiotemporal convolutional block is used to extract the spatiotemporal correlation of data, and the fully connected output layer is used for the final result prediction. The graph structure data constructed by the present invention has the target power station and the spatiotemporally associated power station as nodes, the graph structure data as input, and the irradiance of the target power station as output. The unified prediction model is obtained by training the model, which is represented by GNN.

[0041] The present invention evaluates the constructed prediction model through two methods: prediction accuracy and comparison with a benchmark prediction model. It not only confirms that the prediction accuracy of the constructed unified prediction model is high, but also under different weather conditions, the present invention can effectively overcome the problem that the traditional irradiance prediction model cannot quickly fit the instantaneous changes for sudden cloud cover events. Regardless of whether it is a sudden or non-sudden meteorological condition, it helps to improve the prediction accuracy of ultra-short-term irradiance. The evaluation results show that the prediction results of the unified prediction model of the present invention are very close to the actual irradiance situation. It can be seen that the present invention improves the accuracy of ultra-short-term irradiance prediction, which is conducive to further building a more accurate short-term power prediction model for renewable energy power generation. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 This is a schematic diagram of irradiance using the interpolation technology based on the clear sky index of the present invention.

[0043] Figure 2 This is a schematic diagram of the spatiotemporal graph convolutional network framework flow of the present invention.

[0044] Figure 3 This is the ultra-short-term prediction effect diagram for the first point.

[0045] Figure 4 This is the ultra-short-term prediction effect diagram for the 4th point.

[0046] Figure 5 This is the ultra-short-term prediction effect diagram for the 8th point.

[0047] Figure 6 This is the ultra-short-term prediction effect diagram for the 16th point.

[0048] Figure 7 Comparison of irradiance prediction results under clear sky conditions.

[0049] Figure 8 Comparison of irradiance prediction results under cloudy weather conditions. DETAILED DESCRIPTION

[0050] The technical solutions in the embodiments of the present invention are described clearly and completely below. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0051] The following are specific embodiments:

[0052] S1. Preprocessing of the surface irradiance data around the new energy power station from the measured historical data and satellite multi-channel data inversion:

[0053] S1-1. Preprocessing of historical irradiance measurement data includes filling in missing data and identifying, deleting and filling in abnormal data:

[0054] S1-1-1. The missing data in the measured historical data are filled in by interpolation technology based on the clear sky index:

[0055] The present invention uses satellite inverted irradiance data to calculate the clear sky index in the timestamp of missing data, and then multiplies the clear sky index in the timestamp by the expected irradiance in clear weather to fill the gap in the missing timestamp. The result is shown in the attached Figure 1 .

[0056] S1-1-2. For abnormal data in the measured historical data, we identify and delete them based on the Laida criterion, i.e. the 3-sigma criterion, and use the KNN method to fill in the missing values:

[0057] The present invention identifies abnormal data based on the 3-sigma criterion, and its principle is shown in Formula 3 and Formula 4:

[0058]

[0059] In Formula 3, x is the new energy data matrix to be identified containing various features, is the matrix containing the mean values ​​of each feature, and σ is the matrix containing the standard deviations of each feature;

[0060] In Formula 4, i is the feature dimension of the data to be identified, and m is the number of data samples.

[0061] Based on the 3-sigma criterion under the normal distribution, if the difference between each eigenvalue of the data and its mean value exceeds 3 times the standard deviation, it can be regarded as abnormal data.

[0062] In the present invention, the new energy raw data that is outside the average value by 3σ is considered as abnormal data, and the probability is 0.26%.

[0063] The simplest way to deal with the identified abnormal data is to directly delete it. However, the reduction in data volume caused by direct deletion may affect the integrity of the new energy output time series and the effect of model training, so abnormal data repair is required.

[0064] The present invention adopts the KNN method to fill in missing values. The application of the KNN method in the completion of abnormal data is mainly based on the fact that missing values ​​can be represented according to the average value of the nearest k neighbors. For the completion of multivariate sequences, the nearest k neighbors can be selected in high-dimensional space for implementation.

[0065] S1-2. Preprocessing of surface irradiance data around new energy power stations based on satellite multi-channel data inversion:

[0066] S1-2-1. Establish an irradiance deviation correction model and calculate the correction factor for overlapping data:

[0067] Methods based on satellite-derived solar irradiance may be subject to systematic errors, which may result in an overall over- or underestimation trend, called bias, characterized by a mean bias statistic. When high-quality ground-measured irradiance is available, the bias in the satellite-derived irradiance can be reduced by calculating correction factors a and b for overlapping data.

[0068] Using the preprocessed measured irradiance data and the surface irradiance data of the area around the new energy power station inverted by satellite multi-channel data, the difference between the fitting straight line and the straight line, that is, y=x, is corrected to obtain the correction factors a and b. The deviation of the new estimated value can be ignored, and an irradiance deviation correction model is established.

[0069] S1-2-2. Use the irradiance deviation correction model to correct the inverted irradiance time series data to obtain the corrected inverted irradiance time series data:

[0070] Substitute the ground-measured target power station irradiance time series data and the inverted irradiance time series data into formula 2 to obtain the corrected inverted irradiance time series data;

[0071] The new estimate is calculated using Equation 2, which is:

[0072] y new =y satellite -[(a-1)x groud +b] Formula 2,

[0073] In formula 2, y new Represents the corrected irradiance data, y satellite represents the surface irradiance data of the area around the new energy power station inverted by satellite multi-channel data, x ground represents the measured irradiance data after preprocessing, and a and b represent correction factors.

[0074] S2. Build a unified prediction model based on spatiotemporal graph convolutional neural network:

[0075] S2-1. Construct a prediction model of spatiotemporal graph convolutional neural network with graph structure data as input and target power station irradiance as output. For the schematic diagram, see Figure 2 :

[0076] In the present invention, the neural network is composed of two spatiotemporal convolution blocks and a fully connected layer. The spatiotemporal convolution blocks are used to extract the spatiotemporal correlation of data. Each spatiotemporal convolution block is formed into a "time-space-time" structure, including an upper time-gated convolution layer, a lower time-gated convolution layer and an intermediate spatial graph convolution layer; an output layer is connected after the two spatiotemporal convolution blocks, that is, the fully connected layer output layer is used for the final result prediction, specifically, the time dimension of the original output data after the two spatiotemporal convolution blocks is merged by using the convolution of the time dimension, and the final prediction data is output after a convolution after the merger. The final prediction data is a graph of the next time dimension;

[0077] Each spatiotemporal convolution block uses a gating mechanism in the time dimension, represented by GLU, to implement a 1D causal convolution. The spatiotemporal convolution block input data p l , output data p according to formula 1 l+1 , Formula 1 is:

[0078]

[0079] In formula 1, and Represent the convolution kernels on two time convolutions, Θ l Represents the convolution kernel on the graph convolution, and ReLU represents the rectified linear unit function.

[0080] S2-2. Use the prediction model training of the spatiotemporal graph convolutional neural network to obtain a unified prediction model:

[0081] The preprocessed data used in this embodiment comes from the Gansu New Energy Power Station, which is about 989 meters above sea level and has an installed capacity of 30MWp. The data set spans from January 1, 2023 to May 31, 2023, with the first five months as the training set and the last month as the test set. Since there is only one measured data station, the surface irradiance data of the area around the new energy power station inverted by satellite multi-channel data is used to achieve spatiotemporal dependency modeling.

[0082] The model input is graph structure data, which is arranged in time series, and the output is the target power station irradiance sequence for the next 4 hours with a time resolution of 15 minutes.

[0083] Set initialization parameters based on the spatiotemporal graph convolutional neural network model and train the model on the training set.

[0084] S3. Use the obtained unified prediction model to predict ultra-short-term irradiance: verify the ultra-short-term irradiance prediction effect of the above spatiotemporal graph convolutional network model in the test set. The ultra-short-term irradiance prediction effects of the 1st point, the 4th point, the 8th point and the 16th point are shown in the attached figure. Figure 3-6 .

[0085] In this embodiment, the prediction evaluation index is the mean absolute error, denoted by MAE, and the root mean square error, denoted by RMSE, which represent the mean value of the absolute error and the root of the mean value of the square of the error, respectively.

[0086] Compared with MAE, RMSE does not treat each error value equally. It gives larger errors a greater weight, which means that an error value that is too large will make the RMSE value too high.

[0087] The average formula for prediction accuracy is:

[0088]

[0089] In Equations 5 and 6, i represents the number of hours of forecast data, l is the time range, M is the total number of forecast data, and I measured is the measured value of surface irradiance, I predicted The irradiance predicted by the model. The smaller the RMSE and MAE values ​​are, the smaller the prediction error is and the more accurate the information provided by the prediction is.

[0090] The prediction accuracy calculation results of this embodiment are shown in Table 1.

[0091] Table 1: Prediction accuracy calculation results

[0092]

[0093] In order to further evaluate the effectiveness of the method of the present invention, the GNN model of the present invention can also be proposed in this embodiment for comparison with the benchmark prediction model. In this embodiment, the LSTM model is selected as the benchmark prediction model. The results are shown in Figure 7 and Figure 8 .

[0094] from Figure 7 and Figure 8 It can be seen that the model prediction effect is not much different under clear sky conditions; under cloudy and other weather conditions, irradiance is prone to mutations. The correlation between new energy data in different spatial locations and the simulation of this spatiotemporal information through the spatiotemporal graph convolutional network can effectively overcome the problem that the traditional irradiance prediction model cannot quickly fit the instantaneous changes of sudden cloud cover events. It helps to improve the prediction accuracy of ultra-short-term irradiance under both mutation and non-mutation meteorological conditions.

Claims

1. A method for predicting ultra-short-term irradiance of a new energy power station considering time and space dependence, characterized in that: The steps are: S1. Preprocessing of surface irradiance data around new energy power stations from measured historical data and satellite multi-channel data inversion; S2. Construct a unified prediction model based on spatiotemporal graph convolutional neural network; The S2 step includes: S2-1. Construct a prediction model of a spatiotemporal graph convolutional neural network with graph structure data taking the target power station and spatiotemporally related power stations as nodes, graph structure data as input, and target power station irradiance as output; S2-2, using the data preprocessed in step S1 to train the prediction model of the spatiotemporal graph convolutional neural network to obtain a unified prediction model; S3. Use the obtained unified prediction model to perform ultra-short-term irradiance prediction.

2. According to the method for predicting ultra-short-term irradiance of a new energy power station considering time and space dependence according to claim 1, it is characterized in that: The spatiotemporal graph convolutional neural network described in step S2 is composed of two spatiotemporal convolution blocks and a fully connected layer. The spatiotemporal convolution blocks are used to extract the spatiotemporal correlation of data. Each spatiotemporal convolution block includes a first time-gated convolution layer located at the upper layer, a second time-gated convolution layer located at the lower layer, and a spatial graph convolution layer located in the middle; the fully connected output layer is used for the final result prediction, specifically, the time dimension of the original output data is merged by convolution of the time dimension, and the final prediction data is output after the merging through a convolution. The prediction data is a graph of the next time dimension.

3. The method for predicting ultra-short-term irradiance of a new energy power station considering time and space dependence according to claim 2 is characterized in that: Each spatiotemporal convolution block uses a 1D causal convolution implemented using a gating mechanism in the time dimension. The spatiotemporal convolution block input data p l , output data p according to formula 1 l+1 , Formula 1 is: In formula 1, and denote the convolution kernels on the first and second temporal convolutions, Θ l Represents the convolution kernel on the graph convolution, and ReLU represents the rectified linear unit function.

4. According to the method for predicting ultra-short-term irradiance of a new energy power station considering time and space dependence in claim 1, it is characterized in that: The pretreatment comprises: S1-1. Preprocessing of historical irradiance measurement data includes filling in missing data and identifying, deleting and filling in abnormal data: The missing data in the measured historical data are filled in by interpolation technology based on the clear sky index; Abnormal data in measured historical data are identified and deleted based on the 3-sigma criterion, and the KNN method is used to fill in missing values; S1-2. The preprocessing of the surface irradiance data of the area around the new energy power station inverted from the satellite multi-channel data is corrected by using the irradiance deviation correction model.

5. The method for predicting ultra-short-term irradiance of a new energy power station considering time and space dependence according to claim 4 is characterized in that: The interpolation technology of the clear sky index is to use the irradiance data inverted by satellite to calculate the clear sky index within the timestamp of the missing data, and then fill the gaps in the missing timestamp by multiplying the clear sky index in the timestamp with the expected irradiance in clear weather.

6. The method for predicting ultra-short-term irradiance of a new energy power station considering time and space dependence according to claim 1 is characterized in that: The correction using the irradiance deviation correction model includes: S1-2-1. Establish an irradiance deviation correction model and calculate the correction factor for overlapping data; S1-2-2. Use the irradiance deviation correction model to correct each inverted irradiance time series data to obtain corrected inverted irradiance time series data.

7. The method for predicting ultra-short-term irradiance of a new energy power station considering time and space dependence according to claim 6 is characterized in that: Step S1-2-1 uses the pre-processed measured irradiance data and the surface irradiance data of the area around the new energy power station inverted by satellite multi-channel data to correct the difference between the fitting straight line and the straight line, and calculate the correction factors a and b; Step S1-2-2 calculates the new estimated value through equation 2, which is: y new = y satellite -[(a - 1)x ground + b] Equation 2 In formula 2, y new Represents the corrected irradiance data, y satellite represents the surface irradiance data of the area around the new energy power station inverted by satellite multi-channel data, x ground represents the measured irradiance data after preprocessing, and a and b represent correction factors.

Images