A short-term fine-grained photovoltaic power plant power probability prediction method

Abstract

The present invention proposes a short-term fine-grained photovoltaic power station power probability prediction method, comprising the following steps: 1) photovoltaic power station power and meteorological data acquisition, and missing value and abnormal value processing; 2) extracting meteorological features, and performing feature selection on meteorological features; 3) constructing feature vectors based on time features and meteorological features; 4) constructing a short-term photovoltaic power station power probability prediction model based on the DeepAR model and the feature vector, and obtaining the confidence interval and prediction value of the prediction point. The present invention can be applied to fine-grained photovoltaic power station power probability prediction, and can output a prediction result with statistical interpretability. Not only is the prediction effect good, but also the confidence intervals of different confidence levels of the prediction point can be obtained.

Description

Technical Field

[0001] The invention relates to the technical field of photovoltaic power station power generation prediction. Technical Background

[0002] As an important way to develop and utilize solar energy, photovoltaic power generation has developed rapidly due to its advantages of no fuel consumption, no pollutant emissions, flexible application forms, unlimited capacity scale, safety and reliability, and simple maintenance. Photovoltaic power generation has obvious intermittent and random fluctuation characteristics, which restricts the safe and reliable consumption of photovoltaic power generation. Improving the accuracy of power generation prediction is conducive to improving the utilization rate of photovoltaic power generation equipment.

[0003] Photovoltaic power generation prediction can use power generation curves or meteorological factors to calculate similar days, and use the nearest similar day to predict power generation. With the development of artificial intelligence technology, the most effective method is to use machine learning models, take meteorological forecast data and the historical photovoltaic power generation of photovoltaic power stations as features, output future photovoltaic power generation values, and realize the prediction function. Existing photovoltaic power generation predictions mostly focus on the prediction of power generation values ​​at a certain point in time. The amount of information contained in the point prediction is very small, and the existence of prediction errors is inevitable, which cannot reflect the volatility of the prediction results under different meteorological environments. Summary of the invention

[0004] The present invention aims to address the problems and shortcomings of the above-mentioned prior art and provide a short-term fine-grained photovoltaic power station power probability prediction method, which predicts the fine-grained photovoltaic power station power generation in the next N hours based on the historical power generation data of the photovoltaic power station and future weather forecast data.

[0005] The technical solution adopted by the present invention is: a short-term fine-grained photovoltaic power plant power probability prediction method, comprising the following steps:

[0006] Step 1: Data acquisition and preprocessing.

[0007] Obtain historical meteorological data of the photovoltaic power station at one-hour sampling intervals and power generation data at 15-minute intervals. The meteorological data includes measured irradiance, temperature, weather type, wind speed, wind direction, humidity, pressure, and precipitation.

[0008] For each data field, values ​​outside the mean plus or minus 3 times the standard deviation are set as outliers.

[0009] For outliers and missing values, directly replace them with the last valid value.

[0010] Step 2: Meteorological feature extraction and meteorological feature selection. For time t h, irradiance, temperature, wind speed, wind direction, humidity, and pressure are included in the meteorological feature set. At the same time, considering the cleaning effect of rain on dust, precipitation is included in the meteorological feature set. Calculate the clear sky index:

[0011]

[0012] Where I(t h ) is the irradiance, I sc is the solar constant, d is the date, δ is the solar declination angle, Latitude of photovoltaic power station, θ t is the solar hour angle.

[0013] By using the combination of existing meteorological features, its high-order characteristic meteorological features are calculated, that is, the second-order interaction terms between the meteorological feature sets.

[0014] In order to avoid too many meteorological features, the XGboost model is used to h Meteorological characteristics, high-order meteorological characteristics, weather type, clear sky index and photovoltaic power station time t h The relationship between the measured irradiance and the model is fitted and trained, the importance of each feature is calculated for feature selection, and the model fitting value is used as the corrected irradiance.

[0015] Finally, weather type, clear sky index, corrected irradiance, irradiance, temperature, weather type, wind speed, wind direction, humidity, pressure, precipitation and more important high-order characteristic meteorological features were selected as the meteorological feature set T nwp .

[0016] Step 3: Construct feature vectors based on time characteristics and meteorological characteristics.

[0017] Construct a time feature set, considering the month, day of the year, day of the month, day of the week, hour, and minute of time t to form the time feature set T of time t time,t .

[0018] For time t, consider the whole point t h Meteorological characteristics of T nwp，th , distance from the whole point t h The time distance, the next full point t h+1 Meteorological characteristics Distance to next full point t h+1 The time distance constitutes the meteorological characteristics T at time t nwp,t .

[0019] Meteorological characteristics at time t nwp,t and the time feature set T time,t Construct the feature vector T t .

[0020] For the eigenvector T t The numerical features are numerically normalized. In this invention, the feature vector T t The numerical features are scaled to the interval [-0.5, 0.5], and the feature vector T t The category features are encoded one-hot to obtain the final feature vector X t , the observed value is the photovoltaic power generation z at time t t .

[0021] Step 4: Construct a short-term photovoltaic power station power probability prediction model based on the DeepAR model and feature vector to obtain the confidence interval and prediction value of the prediction point.

[0022] The prediction model is divided into two stages: training process and prediction process:

[0023] Training process: During training, at each time step t, the network input feature is recorded as x i,t , the value of the previous time step is recorded as z i,t-1 , and the state h of the previous time step i,t-1 , calculate the current state Then calculate the parameters of the likelihood l(z|θ) Finally, maximize the log-likelihood To learn the parameters of the network. The specific form of depends on the likelihood function l(z|θ), which needs to be selected based on the statistical characteristics of the data itself. If Gaussian distribution is selected, θ=(μ,σ), and the likelihood function is

[0024]

[0025] During training, the training time window length is set to one day, that is, the time window of the neuron LSTM. The prediction length is 4*n points, that is, the prediction value of every 15 minutes in the next n hours is obtained. t and photovoltaic power generation power z t Construct a data set D = (X, z). Divide the data set D into training set D according to the ratio of 8:1:1. train , validation set D val , test set D test In the training set D train and validation set D val The prediction model is trained on the validation set D val Responsible for evaluating the fitting loss during training, providing optimization direction for training, and preventing overfitting.

[0026] Prediction process: After training is completed, historical data of t<t0 is input into the network to obtain the initial state Use the ancestor sampling method to obtain the result: for t0, t0+1, K, T, randomly sample z at each time step~ i,t ~l(·|θ i,t ), this sample is used as the input for the next time step. Repeating this process can obtain a series of sample values ​​from t0 to T, which can be used to calculate the required target values, such as quantiles, expectations, etc. The expectation of the sample value is used as the predicted value, and the confidence interval of the predicted value is calculated according to the set confidence level.

[0027] The implementation of the present invention has the following beneficial effects:

[0028] The present invention can increase the joint characterization of photovoltaic power generation through meteorological characteristics and time characteristics within the time window, thereby enhancing the prediction accuracy of the model; convert the photovoltaic power generation value prediction into the prediction of the probability distribution of the prediction point, obtain the confidence interval and prediction value corresponding to the target confidence level through sampling, and better explain the prediction results; and can perform short-term fine-grained power generation prediction, providing a basis for photovoltaic power station power generation tracking and grid-connected scheduling. The present invention not only has a good prediction effect, but also can obtain confidence intervals of different confidence levels of the prediction point. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 It is an overall schematic diagram of a short-term fine-grained photovoltaic power station power probability prediction method of the present invention.

[0030] Figure 2 This is a schematic diagram of the model training process.

[0031] Figure 3 It is a schematic diagram of the model prediction process.

[0032] Figure 4 It is a schematic diagram of the prediction results. DETAILED DESCRIPTION

[0033] The technical solution of the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0034] Figure 1 The present invention shows that a short-term fine-grained photovoltaic power plant power probability prediction method is provided, comprising the following steps:

[0035] Step 1, data acquisition and preprocessing;

[0036] Step 2, meteorological feature extraction and meteorological feature selection;

[0037] Step 3, constructing a feature vector based on time features and meteorological features;

[0038] Step 4: construct a short-term photovoltaic power station power probability prediction model based on the DeepAR model and feature vector to obtain the confidence interval and prediction value of the prediction point.

[0039] The above steps are described in detail as follows:

[0040] Step 1: Data acquisition and preprocessing. Obtain historical weather forecast data, measured irradiance data, and power generation data at 15-minute intervals for the photovoltaic power station with a sampling interval of one hour. The meteorological data obtained include measured irradiance, temperature, weather type, wind speed, wind direction, humidity, pressure, and precipitation.

[0041] For each data field, values ​​outside the mean plus or minus 3 times the standard deviation are considered outliers.

[0042] For abnormal values ​​and missing values, since each data field has a time series relationship, the last valid value is used directly instead.

[0043] Step 2: Meteorological feature extraction and meteorological feature selection. For time t h , irradiance, temperature, wind speed, wind direction, humidity, and pressure are included in the meteorological feature set. At the same time, considering the cleaning effect of rain on dust, precipitation is included in the meteorological feature set. Calculate the clear sky index

[0044]

[0045] Where I(t h ) is the irradiance, I sc is the solar constant, d is the date, δ is the solar declination angle, Latitude of photovoltaic power station, θ t is the solar hour angle.

[0046] By using the combination of existing meteorological features, its high-order characteristic meteorological features, that is, the second-order interaction terms between the meteorological feature sets, are calculated. Since the environmental factors in the meteorological feature set affect each other, the high-order characteristic meteorological features can provide better prediction results.

[0047] In order to avoid too many meteorological features, the XGboost model is used to h Meteorological characteristics, high-order meteorological characteristics, weather type, clear sky index and photovoltaic power station time t h The XGboost model can calculate the importance of each feature for feature selection and use the model fitting value as the corrected irradiance.

[0048] Finally, weather type, clear sky index, corrected irradiance, irradiance, temperature, weather type, wind speed, wind direction, humidity, pressure, precipitation and more important high-order characteristic meteorological features were selected as the meteorological feature set T nwp .

[0049] Step 3: Construct feature vectors based on time features and meteorological features: time features, power generation, meteorological features of the previous hour, time distance from the previous hour, meteorological features of the next hour, and time distance from the next hour.

[0050] Construct a time feature set, considering the month, day of the year, day of the month, day of the week, hour, and minute of time t to form the time feature set T of time t time,t .

[0051] Since there is no 15-minute fine-grained meteorological characteristic data, for time t, consider the hour t h Meteorological characteristics Distance from the whole point t h The time distance, the next full point t h+1 Meteorological characteristics Distance to next full point t h+1 The time distance constitutes the meteorological characteristics T at time t nwp,t .

[0052] Meteorological characteristics at time t nwp,t and the time feature set T time,t Construct the feature vector T t .

[0053] For the eigenvector T t The numerical features are numerically normalized. In this invention, the feature vector T t The numerical features are scaled to the interval [-0.5, 0.5], and the feature vector T t The category features are encoded one-hot to obtain the final feature vector X t , the observed value is the photovoltaic power generation z at time t t .

[0054] Step 4: Construct a short-term photovoltaic power station power probability prediction model based on the DeepAR model and feature vector to obtain the confidence interval and prediction value of the prediction point.

[0055] The main framework of the model adopts the probabilistic neural network DeepAR model, whose basic network structure is a memory recurrent neural network. The neurons use LSTM to learn the conditional probability distribution parameters of future time period sequence values ​​based on historical period sequence values.

[0056] The prediction model is divided into two stages:

[0057] The training process is as follows Figure 2 As shown, during training, at each time step t, the network input feature is recorded as x i,t , the value of the previous time step is recorded as z i,t-1 , and the state h of the previous time step i,t-1 , calculate the current state Then calculate the parameters of the likelihood l(z|θ) Finally, maximize the log-likelihood To learn the parameters of the network. The specific form of depends on the likelihood function l(z|θ), which needs to be selected based on the statistical characteristics of the data itself. If Gaussian distribution is selected, θ=(μ,σ), and the likelihood function is:

[0058]

[0059] During training, the training time window length is set to one day, that is, the time window of the neuron LSTM. The prediction length is 4*n points, that is, the prediction value of every 15 minutes in the next n hours is obtained. t and photovoltaic power generation power z t Construct a data set D = (X, z). Divide the data set D into training set D according to the ratio of 8:1:1. train , validation set D val , test set D test In the training set D train and validation set D val The prediction model is trained on the validation set D val Responsible for evaluating the fitting loss during training, providing optimization direction for training, and preventing overfitting.

[0060] The prediction process is as follows Figure 3 As shown, after the training is completed, the historical data of t<t0 is input into the network to obtain the initial state Use the ancestor sampling method to obtain the results: for t0, t0+1, K, T, randomly sample at each time step This sample is used as the input of the next time step. Repeating this process can obtain a series of sample values ​​from t0 to T, and these sample values ​​can be used to calculate the required target values, such as quantiles, expectations, etc. In the present invention, the expectation of the sample value is used as the predicted value, and the confidence interval of the predicted value is calculated according to the set confidence level.

[0061] Specifically, in the present invention, the target prediction length is set to 16 points, and the prediction results of every 15 minutes in the next 4 hours are obtained. The confidence levels are set to 80% and 40% respectively, and the confidence intervals corresponding to different confidence levels are obtained. The mean of the sample values ​​of each prediction point is calculated as the prediction value. The prediction results are as follows: Figure 4 As shown, when the prediction starting point is 2018.02.179, the darkest black line is the true value, and the other gray areas are the predicted values ​​and confidence intervals corresponding to different confidence levels. It can be seen that the farther the prediction point is from the prediction starting point, the wider the confidence interval, which means the greater the uncertainty.

[0062] The photovoltaic power generation prediction model is established according to the different prediction time settings. The longer the prediction time, the lower the accuracy, and the shorter the prediction time, the higher the accuracy. The prediction results of different models are statistically shown in the following table:

[0063] Prediction length Prediction granularity MSE MAPE 4h 15 minutes 1.877 0.091 15 minutes 15 minutes 0.306 0.051

[0064] Obviously, the above embodiments are merely examples for clearly illustrating the technical solutions of the present invention, and are not intended to limit the implementation methods. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the implementation methods here. The obvious changes or modifications derived therefrom are still within the protection scope of the present invention.

Claims

1. A short-term fine-grained photovoltaic power plant power probability prediction method, characterized in that: The following steps are involved: Step 1, data acquisition and preprocessing; Step 2, meteorological feature extraction and meteorological feature selection; Step 3, constructing a feature vector based on time features and meteorological features; Step 4: construct a short-term photovoltaic power station power probability prediction model based on the DeepAR model and feature vector to obtain the confidence interval and prediction value of the prediction point; The step 2 comprises: (1) For time t h , put the meteorological data into the meteorological feature set and calculate the clear sky index: Where I(t h ) is the irradiance, I sc is the solar constant, d is the date, δ is the solar declination angle, is the latitude of the photovoltaic power station, θ t is the solar hour angle; (2) Using the combination of existing meteorological features, calculate its high-order characteristic meteorological features, that is, the second-order interaction terms between the meteorological feature sets; (3) Use the XGboost model to analyze the time t h Meteorological characteristics, high-order meteorological characteristics, weather type, clear sky index and photovoltaic power station time t h The relationship between the measured irradiance and the model is trained to fit the model, the importance of each feature is calculated to select the features, and the model fitting value is used as the corrected irradiance. (4) Finally, weather type, clear sky index, corrected irradiance, irradiance, temperature, wind speed, wind direction, humidity, pressure, precipitation and high-order characteristic meteorological features are selected as the meteorological feature set T nwp .

2. A short-term fine-grained photovoltaic power plant power probability prediction method according to claim 1, characterized in that: In step 1, the acquired data includes historical meteorological data of the photovoltaic power station at one-hour intervals and power generation data at 15-minute intervals; the meteorological data includes measured irradiance, temperature, weather type, wind speed, wind direction, humidity, pressure, and precipitation; The preprocessing includes preprocessing abnormal values ​​and missing values, specifically using the last valid value to replace the abnormal value or missing value; the abnormal value is a value outside the mean value plus or minus 3 times the standard deviation.

3. A short-term fine-grained photovoltaic power plant power probability prediction method according to claim 1, characterized in that: The step 3 comprises: (1) Construct a time feature set, considering the month, day of the year, day of the month, day of the week, hour, and minute at time t to form the time feature set T at time t time,t ; (2) For time t, consider the whole point t h Meteorological characteristics Distance from the whole point t h The time distance, the next full point t h+1 Meteorological characteristics Distance to next full point t h+1 The time distance constitutes the meteorological characteristics T at time t nwp,t ; (3) Meteorological characteristics T at time t nwp,t and the time feature set T time,t Construct the feature vector T t ; (4) For the eigenvector T t The numerical features in are numerically normalized, and the feature vector T t The numerical features are scaled to the interval [-0.5, 0.5], and the feature vector T t The category features are encoded one-hot to obtain the final feature vector X t , the observed value is the photovoltaic power generation z at time t t .

4. A short-term fine-grained photovoltaic power plant power probability prediction method according to claim 1, characterized in that: In step 4, the prediction model is divided into two stages: training process and prediction process: (1) Training process: During training, at each time step t, the network input feature is recorded as x i,t , the value of the previous time step is recorded as z i,t-1 , and the state h of the previous time step i,t-1 , calculate the current state Then calculate the parameters of the likelihood l(z|θ) Finally, maximize the log-likelihood To learn the parameters of the network; The specific form of depends on the likelihood function l(z|θ), which is selected based on the statistical characteristics of the data itself. If Gaussian distribution is selected, θ=(μ,σ), and the likelihood function is: During training, the training time window length is set to one day, that is, the time window of the neuron LSTM; the prediction length is 4*n points, that is, the prediction value of every 15 minutes in the next n hours is obtained; the feature vector X of all time t is set to t and photovoltaic power generation power z t Construct a data set D = (X, z); divide the data set D into training set D according to the ratio of 8:1:1 train , validation set D val , test set D test ; In the training set D train and validation set D val The prediction model is trained on the validation set D val Responsible for evaluating the fitting loss during training, providing optimization direction for training, and preventing overfitting; (2) Prediction process: After training is completed, historical data with t<t0 is input into the network to obtain the initial state Use the ancestor sampling method to obtain the results: for t0, t0+1, K, T, randomly sample at each time step This sample is used as the input for the next time step; repeating this process can obtain a series of sample values ​​from t0 to T, and these sample values ​​are used to calculate the required target value.

Images