Photovoltaic output prediction method considering haze / dust factor

Abstract

The invention provides a photovoltaic output prediction method considering haze / dust factors. The method comprises the following steps: firstly, preprocessing data, screening out abnormal data and invalid data, and complementing missing data; the method comprises the following steps: firstly, analyzing haze / sand dust data, then classifying weather data by using a decision tree algorithm, then deeply understanding an internal rule and an operation mechanism by analyzing a haze / sand dust formation mechanism and a photovoltaic power generation principle, determining main characteristics, and then firstly establishing a haze / sand dust influence model 1 and a photovoltaic power generation power prediction model 2 for the haze / sand dust data; the model 1 obtains influence parameters of haze / dust factors on the photovoltaic power generation power, and the model 2 outputs a predicted value of the photovoltaic power generation power. The advantages of mechanism analysis and a data driving model are combined, and the accuracy and reliability of the model are improved.

Description

Technical Field

[0001] The present invention belongs to photovoltaic output prediction technology, and relates to a photovoltaic output prediction method taking haze / sand dust factors into consideration. Background Art

[0002] Photovoltaic power generation is a source of new energy for the power system and an effective way to reduce carbon emissions. However, photovoltaic power generation is affected by weather factors and the power generation is fluctuating. In order to balance the power generation side and the power consumption side, it is necessary to predict the power generation of photovoltaic power stations in a certain period of time in the future. Research on photovoltaic power station power prediction technology is an important means to improve the accuracy of new energy prediction, ensure the safe and reliable operation of the power grid, and improve the overall benefits of power companies. Typical problems currently existing in the prediction of new energy power in some areas are: first, the numerical weather forecast is empty and the prediction deviation is large; second, the perception ability of extreme and abnormal weather prediction is weak; third, the quality of basic data is worrying, which seriously affects the accuracy of prediction.

[0003] In some areas rich in photovoltaic resources, large-scale haze and sandstorms frequently occur, which has a serious impact on photovoltaic power generation output. The causes and changes of large-scale haze / sandstorms are complex, and the uncertainty of the impact on solar radiation has affected the output forecast accuracy of photovoltaic stations. The current photovoltaic output power prediction technology is based on historical actual power, numerical weather forecasts, and ground-based cloud maps. It uses statistical analysis methods to calculate future power and predict the output power in the next 2~72 hours. When the photovoltaic output fluctuates slightly, the above method has a certain accuracy, but it cannot provide accurate output prediction results for situations where photovoltaic power generation power drops over a large area due to extreme weather such as haze / sandstorms. Therefore, under possible sandstorm conditions, the accurate prediction of photovoltaic power station output is an urgent problem to be solved. Summary of the invention

[0004] The purpose of the present invention is to propose a photovoltaic output prediction method taking haze / sand dust factors into consideration.

[0005] The technical solution of the present invention is a photovoltaic output prediction method considering haze / sand dust factors, the steps are as follows: S1, collect historical data of the target PV power station, including meteorological data and power data, pre-process the data, use linear interpolation to fill in missing data, and process outliers in the data; then normalize the data to eliminate the impact of dimension; S2, using the decision tree algorithm to classify historical data, using the information gain rate to select features, building a decision tree based on the features, and dividing the historical data into three categories: haze / sandstorm days, rainy days turning to sunny days, and other days; S3, by analyzing the formation mechanism of haze / dust and the principle of photovoltaic power generation, determine the main characteristics that affect photovoltaic power generation under the condition of haze / dust factors; S4, for haze / sand and dust data, based on mechanism analysis, a neural network algorithm is used to establish a haze / sand and dust impact model 1; S5, for haze / sand dust data, based on mechanism analysis, a linear regression algorithm is used to establish a photovoltaic power generation prediction model 2; S6, using the cross-validation method, in each iteration of cross-validation, the model is trained using the training set, and then the model is evaluated using the test set. The results of all cross-validation iterations are summarized and analyzed, and the average performance index and standard deviation are calculated to evaluate the stability and reliability of the model, and finally a suitable model is obtained; S7, taking the features determined in S3 as the input of Model 1 and Model 2, and realizing the photovoltaic output prediction considering the haze / dust factors through the two outputs of the two models.

[0006] The formula for normalizing the historical data in step S1 is: ,in, L’ represents the normalized data; L represents the data before normalization, L min Indicates the minimum value among all the data. L max Indicates the maximum value in the data.

[0007] The decision tree algorithm in step S2 is C4.5, and the decision tree is constructed by: assuming random variables Y is a discrete random variable that takes a finite number of values, and its probability distribution p i is: P{Y=y i = p i , i =1,2,…,n, random variable Y Entropy H ( Y )The formula is: , i =1,2,…,n, let the training data set be D, |D| represents the number of samples, and there are K classes C k ,i=1,2,…,K,|C k | belongs to class C k The number of samples, the entropy of the training data set D H ( D )The formula is: ,random variable Y Make class C kThe degree of uncertainty reduction, that is, the change in system entropy, is the information gain Gain ( D , Y ), the larger the information gain, the more suitable the variable feature is for classification. Y The information gain rate is the ratio of its information gain to entropy. According to the above formula, (1) first calculate the original entropy of the data set, (2) calculate the entropy of each variable feature, (3) calculate the information gain rate of each variable feature, and select the feature with the largest information gain rate as the root node, (4) treat each node as a new tree, select the remaining features, repeat the above steps, continue to find branch nodes through the information gain rate, and finally generate a decision tree. Entropy is a measure of the uncertainty of a random variable.

[0008] A part of the historical data is used to form the training data set D. Wind speed, visibility, PM2.5, PM10, cloud cover, and ambient temperature are used as random variables. The above-mentioned method of constructing a decision tree is used to divide the historical data into three categories: haze / sandstorm days, rain-turned-sunny days, and other days.

[0009] In the step S3, according to the formation mechanism of haze / dust, wind speed, visibility, PM2.5, PM10, and ambient temperature are determined as key factors to characterize haze / dust. Wind speed, visibility, PM2.5, PM10, ambient temperature, DNI, and component cleanliness are in a nonlinear relationship, and a neural network model is used for prediction. According to the power generation mechanism of the photovoltaic power generation system, DNI and component cleanliness are determined as key factors affecting photovoltaic power generation. Because the data of component cleanliness cannot be directly obtained, the difference between component temperature and ambient temperature is used as a representative quantity of component cleanliness. DNI, component cleanliness, photovoltaic power generation power of the previous day and predicted photovoltaic power generation power are in a linear relationship, and a linear regression algorithm is used to build a model for prediction.

[0010] In step S4, an echo state network ESN is used, and it is set at time t that the input is x ( t ), the input node is m , the reserve pool status is y ( t ), the number of nodes is n , the output is z ( t ), the number of nodes is o , the state of each layer can be expressed as: , T is the matrix transposition symbol. The state equations of each layer in the current stage are shown above, and the state equation of the next stage can be obtained as follows: , in the formula f and f outThey are the activation functions of the model state and output equations respectively. The default activation function is the tanh function. W in is the connection weight from the input layer to the reserve pool, W is the connection weight between the storage pools, W out is the connection weight between the reserve pool and the output layer, W back is the feedback output weight, x ( t +1) is the input of the next moment at time t, y ( t +1) is the state of the reserve pool at the next moment after time t, z ( t +1) is the output of the next moment after time t. From the state equation of the next stage of the reserve pool, it can be seen that the current state at each moment is related to the states of multiple moments, and is related to the state response at the current moment and the state response at the previous stage. At the same time, during the update process, the output will be updated through the feedback matrix W back Read and adjust the information.

[0011] Taking wind speed, visibility, PM2.5, PM10, and ambient temperature as input, DNI and component cleanliness as output, and the above data in the haze / dust day category in historical data as the training data set, the echo state network model, namely the haze / dust impact model 1, is obtained.

[0012] In step S5, the following function is established: , , q is the input feature, h For output, i is the characteristic parameter, e is the error, i T q (i) is the predicted value, r (i) is the true value, and the loss function is: , i Indicates the quantity number involved in the calculation, Q Represents the input feature matrix, and the least squares method is used to solve the feature parameters i : .

[0013] Taking DNI, component cleanliness, and photovoltaic power generation of the previous day as input features, the predicted daily photovoltaic power generation is output, and the above data of haze / sandstorm categories in historical data are used as training data sets to obtain photovoltaic power generation prediction model 2.

[0014] In step S6, the k-fold cross-validation method is used to divide the data set into k subsets, each of which is used as a test set, and the rest is used as a training set each time, cross-validated k times, and the average cross-validation performance index of k times is used as the final result. The mean absolute error and root mean square error are used to measure the prediction performance of the model: Mean absolute error: ; Root Mean Square Error: ; Where: a — the number of prediction results, A is the total number of calculations of the Σ symbol formula; P means (a) , P pred (a) ——Actual values ​​and model predicted values.

[0015] The beneficial effects of the present invention are as follows: 1. According to the formation mechanism of haze / dust, wind speed, visibility, PM2.5, PM10, and ambient temperature are determined as key factors for characterizing haze / dust, rather than looking for key factors through historical data correlation, thereby reducing complexity; wind speed, visibility, PM2.5, PM10, ambient temperature, DNI, and component cleanliness are nonlinear relationships, and the use of a neural network model for prediction is more accurate.

[0016] 2. According to the power generation mechanism of the photovoltaic power generation system, DNI and component cleanliness are determined as the key factors affecting the photovoltaic power generation. The component cleanliness data cannot be directly obtained, and the difference between the component temperature and the ambient temperature is used as a representative quantity of the component cleanliness. The influencing factors are clarified, the prediction process is simplified, and the accuracy of the model prediction is increased. There is a linear relationship between DNI, component cleanliness, the photovoltaic power generation of the previous day and the predicted photovoltaic power generation on the day. The linear regression algorithm is used to build a model for prediction, which is more accurate.

[0017] 3. Use the decision tree algorithm to build a decision tree based on the features and divide the data into three categories: haze / sandstorm days, rainy days, and other days. Select the haze / sandstorm category and train the model on the representative data of the haze / sandstorm category to reduce the amount of calculation and avoid overfitting, improve the generalization ability of the model, and improve the prediction accuracy.

[0018] 4. Combining the advantages of mechanism analysis and data-driven models, it can help deal with uncertainties in complex problems, better understand the sources and impacts of these uncertainties, more accurately reflect the behavior of actual systems, and deal with uncertainties in complex problems, improving the accuracy and reliability of the model.

[0019] 5. The training process of the ESN network model is greatly simplified compared to other neural networks, and the network structure is more certain, which is suitable for photovoltaic power generation prediction. It simplifies the complexity of model training and improves computational efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 It is a flow chart of the present invention. DETAILED DESCRIPTION

[0021] This patent takes into account the differences in photovoltaic power under the influence of haze / dust factors, constructs differentiated photovoltaic power prediction models for haze / dust categories, realizes short-term prediction of photovoltaic power, and improves the accuracy of photovoltaic power prediction under different types of weather.

[0022] A photovoltaic output prediction method considering haze / sand dust factors is as follows: 1. Collect historical data of the target PV power station, including meteorological data and power data, pre-process the data, use linear interpolation to fill in missing data, and process outliers in the data; then normalize the data to eliminate the impact of dimension; The data is preprocessed and the normalization formula is: ,in, L’ represents the normalized data; L represents the data before normalization, L min Indicates the minimum value among all the data. L max Indicates the maximum value in the data.

[0023] 2. Use the decision tree algorithm to classify the data, use the information gain rate to select features, build a decision tree based on the features, and divide the data into three categories: haze / sandstorm days, rainy days turning sunny days, and other days; The decision tree algorithm C4.5 is used to classify the data, the information gain rate is used to select features, and the decision tree is constructed based on the features.

[0024] Specifically: Assume that the random variable Y is a discrete random variable with a finite number of values, and its probability distribution p i is: P{Y=y i =p i , i=1,2,…,n; The formula for the entropy H(Y) of a random variable Y is: i=1,2,…,n; Suppose the training data set is D, |D| represents the number of samples, and there are K classes C k ,i=1,2,…,K,|C k | belongs to class C k The entropy H(D) formula of the training data set D is:

[0025] When adding a Y When the random variable Y Make class C k The degree of uncertainty reduction, that is, the change in system entropy, is the information gain Gain ( D , Y ). Different features often have different information gains. The larger the information gain, the more suitable the feature is for classification.

[0026] Y The information gain rate of is the ratio of its information gain to entropy. Entropy is a measure of the uncertainty of a random variable.

[0027] According to the above formula, (1) first calculate the original entropy of the data set, (2) calculate the entropy of each variable feature, (3) calculate the information gain rate of each variable feature, and select the feature with the largest information gain rate as the root node, (4) treat each node as a new tree, select the remaining features, repeat the above steps, continue to find branch nodes through the information gain rate, and finally generate a decision tree.

[0028] A training data set D is composed of a part of historical data. Six environmental factors (wind speed, visibility, PM2.5, PM10, cloud cover, and ambient temperature) are used as features to construct a decision tree, and the data are divided into three categories: haze / sandstorm days, rain-turned-sunny days, and other days.

[0029] 3. By analyzing the formation mechanism of haze / dust and the principle of photovoltaic power generation, the main characteristics that affect photovoltaic power generation under the condition of haze / dust factors are determined.

[0030] Photovoltaic power generation is related to the following factors: direct solar radiation (DNI), photovoltaic module area, module cleanliness, module photoelectric conversion efficiency, inverter conversion efficiency, and line loss. Among them, the photovoltaic module area can be regarded as a constant. The module photoelectric conversion efficiency, inverter conversion efficiency, and line loss gradually decay with the use of the module and equipment. Direct solar radiation (DNI) changes with weather changes. On the one hand, the cleanliness of the module gradually decreases with the accumulation of time, and on the other hand, it will suddenly change due to certain weather. For example, when it rains, the cleanliness of the module suddenly increases sharply, and when it encounters haze / sandstorm weather, the cleanliness of the module suddenly decreases sharply. Rain turns to clear, and the corresponding component cleanliness after rainfall. Classification, haze / sandstorm day, rain turns to clear, and other days.

[0031] Photovoltaic power generation refers to a direct power generation method that uses the photovoltaic effect principle of the PN junction of a semiconductor electronic device such as a solar cell to effectively absorb solar radiation energy and convert it into electrical energy through a conversion device. Dust on the surface of a photovoltaic module will reduce the solar radiation it receives. Dust will also affect the heat dissipation performance of the photovoltaic module, causing the temperature to rise, further reducing the power generation efficiency. According to the power generation mechanism of the photovoltaic power generation system, DNI and module cleanliness are determined to be the key factors affecting the photovoltaic power generation power. The data of module cleanliness cannot be directly obtained, and the difference between the module temperature and the ambient temperature is used as a representative quantity of the module cleanliness. Clarify the influencing factors and simplify the prediction process. Wind speed, visibility, PM2.5, PM10, ambient temperature, DNI, and module cleanliness are nonlinear relationships. The use of a neural network model for prediction increases the accuracy of the model prediction.

[0032] Haze / dust weather is divided into three categories: floating dust, blowing sand, and sandstorm. Floating dust: sand or soil particles suspended in the atmosphere, a weather phenomenon with a horizontal visibility of less than 10 kilometers; blowing sand: wind blows up dust on the ground, making the air quite turbid, and a weather phenomenon with a horizontal visibility of less than 1-10 kilometers. Sandstorm: strong wind blows up dust on the ground, making the air quite turbid, and a weather phenomenon with a horizontal visibility of less than 1 kilometer. When the horizontal visibility is less than 500 meters, it is a strong sandstorm. The main causes of sand and dust are strong winds, unstable air conditions, and sand and dust materials on the ground. According to the formation mechanism of haze / dust, wind speed, visibility, PM2.5, and PM10 are the key factors affecting DNI and component cleanliness.

[0033] According to the above principles, direct solar radiation (DNI) and component cleanliness are taken as key factors affecting photovoltaic power generation. When establishing the model, it is assumed that the four factors of photovoltaic module area, module photoelectric conversion efficiency, inverter conversion efficiency, and line loss are included in the photovoltaic power generation of the previous day. Taking the photovoltaic power generation of the previous day as input, through model training, the weight factors of the above four factors for predicting the photovoltaic power generation of the day can be obtained. For the two factors of direct solar radiation (DNI) and component cleanliness, a model related to weather factors is established, with weather factors as input, and direct solar radiation (DNI) and component cleanliness as output. There is a linear relationship between DNI, component cleanliness, photovoltaic power generation of the previous day and the predicted photovoltaic power generation of the day. The linear regression algorithm is used to build a model for prediction, which is more accurate.

[0034] 4. For haze / sand and dust data, based on mechanism analysis, a neural network algorithm is used to establish a haze / sand and dust impact model 1; Using the neural network algorithm, the haze / dust impact model 1 is established. The specific process is as follows: Echo State Networks (ESN) have the same input and output layers as normal neural network structures, except that in the middle part, the traditional neural network is a hidden layer, while the ESN model is a reserve pool. The units in the reserve pool can be connected to each other or self-connected, and the reserve pool weights are automatically generated. Therefore, during the training of the ESN network model, only the output weights connecting the reserve pool and the output layer need to be trained, which greatly simplifies the complexity of model training.

[0035] Set at time t, the input is x ( t ), the input node is m , the reserve pool status is y ( t ), the number of nodes is n , the output is z ( t ), the number of nodes is o The state of each layer can be expressed as: (1), T is the matrix transpose symbol; The state equations of each layer in the current stage are shown in formula (1), and the state equation of the next stage can be obtained as follows: (2), in formula (2), f and f out They are the activation functions of the model state and output equations respectively. The default activation function is the tanh function. Win is the connection weight from the input layer to the reserve pool, W is the connection weight between the storage pools, W out is the connection weight between the reserve pool and the output layer, W back is the feedback output weight, x ( t +1) is the input of the next moment at time t, y ( t +1) is the state of the reserve pool at the next moment after time t, z ( t +1) is the output of the next moment at time t. From the state update equation (2) of the reserve pool, it can be seen that the current state at each moment is related to the states of multiple moments, and is related to the state response at the current moment and the state response at the previous stage. At the same time, during the update process, the output will be updated through the feedback matrix W back Reading and adjusting information is in line with the memory characteristics of the network.

[0036] In formula (2), W in and W It is initially randomly generated and can be obtained through training W out .

[0037] Taking wind speed, visibility, PM2.5, PM10, and ambient temperature as input, DNI and component cleanliness as output, and the above data in the haze / dust day category in historical data as the training data set, the echo state network model, namely the haze / dust impact model 1, is obtained.

[0038] S5, for haze / sand dust data, based on mechanism analysis, a linear regression algorithm is used to establish a photovoltaic power generation prediction model 2; The linear regression algorithm is used to establish photovoltaic power generation prediction model 2. The specific process is as follows: Create the following function: , ; q is the input feature, h For output, i is the characteristic parameter, e is the error, i T q (i) is the predicted value, r (i) is the true value.

[0039] The loss function is: , the least squares method is used to solve the characteristic parameters i : , Q Represents the input feature matrix.

[0040] Taking DNI, component cleanliness, and photovoltaic power generation of the previous day as input features, the predicted daily photovoltaic power generation is output, and the above data of haze / sandstorm categories in historical data are used as training data sets to obtain photovoltaic power generation prediction model 2.

[0041] 6. Use the cross-validation method. In each iteration of cross-validation, use the training set to train the two models, and then use the test set to evaluate the two models. Summarize and analyze the results of all cross-validation iterations, calculate the average performance index and statistics such as standard deviation to evaluate the stability and reliability of the model. Select appropriate model parameters to avoid overfitting or underfitting and improve the generalization ability of the model.

[0042] The k-fold cross validation method is used to divide the data set into k subsets. Each subset is used as a test set, and the rest are used as training sets each time. Cross validation is performed k times, and the average cross validation performance index of k times is used as the final result. The mean absolute error and root mean square error are used to measure the prediction performance of the two models.

[0043] Mean absolute error: ; Root Mean Square Error: ; Where: a — the number of prediction results, A is the total number of calculations of the Σ symbol formula; P means (a)、P pred (a) ——Actual values ​​and model predicted values.

[0044] Finally, the photovoltaic power generation power prediction considering haze / sand dust factors is realized. Wind speed, visibility, PM2.5, PM10, and ambient temperature are input into model 1 to obtain DNI and component cleanliness; DNI, component cleanliness, and photovoltaic power generation power of the previous day are then input into model 2 to obtain the predicted photovoltaic power generation power of the day.

Claims

1. A photovoltaic output prediction method considering haze / sand dust factors, characterized in that Here are the steps: S1, collect historical data of the target PV power station, including meteorological data and power data, pre-process the data, use linear interpolation to fill in missing data, and process outliers in the data; then normalize the data to eliminate the impact of dimension; S2, using the decision tree algorithm to classify historical data, using the information gain rate to select features, building a decision tree based on the features, and dividing the historical data into three categories: haze / sandstorm days, rainy days turning to sunny days, and other days; S3, by analyzing the formation mechanism of haze / dust and the principle of photovoltaic power generation, determine the main characteristics that affect photovoltaic power generation under the condition of haze / dust factors; S4, for haze / sand and dust data, based on mechanism analysis, a neural network algorithm is used to establish a haze / sand and dust impact model 1; S5, for haze / sand dust data, based on mechanism analysis, a linear regression algorithm is used to establish a photovoltaic power generation prediction model 2; S6, using the cross-validation method, in each iteration of cross-validation, the model is trained using the training set, and then the model is evaluated using the test set. The results of all cross-validation iterations are summarized and analyzed, and the average performance index and standard deviation are calculated to evaluate the stability and reliability of the model, and finally a suitable model is obtained; S7, taking the features determined in S3 as the input of Model 1 and Model 2, and realizing the photovoltaic output prediction considering the haze / dust factors through the two outputs of the two models.

2. A photovoltaic output prediction method considering haze / sand dust factors according to claim 1, characterized in that: The formula for normalizing historical data in step S1 is: ,in, L’ represents the normalized data; L represents the data before normalization, L min Indicates the minimum value among all the data. L max Indicates the maximum value in the data.

3. A photovoltaic output prediction method considering haze / sand dust factors according to claim 1, characterized in that: Step S2: The decision tree algorithm is C4.

5. The decision tree construction method is: let the random variable Y is a discrete random variable that takes a finite number of values, and its probability distribution p i is: P{Y=y i = p i , i =1,2,…,n, random variable Y Entropy H ( Y )The formula is: , i =1,2,…,n, let the training data set be D, |D| represents the number of samples, and there are K classes C k ,i=1,2,…,K,|C k | belongs to class C k The number of samples, the entropy of the training data set D H ( D )The formula is: , the change in system entropy is the information gain Gain ( D , Y ), Y The information gain rate is the ratio of its information gain to entropy. According to the above formula, (1) first calculate the original entropy of the data set, (2) calculate the entropy of each variable feature, (3) calculate the information gain rate of each variable feature, and select the feature with the largest information gain rate as the root node, (4) treat each node as a new tree, select the remaining features, repeat the above steps, continue to find branch nodes through the information gain rate, and finally generate a decision tree; A part of the historical data is used to form the training data set D. Wind speed, visibility, PM2.5, PM10, cloud cover, and ambient temperature are used as random variables. The above-mentioned method of constructing a decision tree is used to divide the historical data into three categories: haze / sandstorm days, rain-turned-sunny days, and other days.

4. The photovoltaic output prediction method considering haze / sand dust factors according to claim 1 is characterized in that: Step S3, according to the formation mechanism of haze / dust, determine wind speed, visibility, PM2.5, PM10, and ambient temperature as key factors to characterize haze / dust. Wind speed, visibility, PM2.5, PM10, ambient temperature, DNI, and component cleanliness are nonlinearly related, and a neural network model is used for prediction. According to the power generation mechanism of the photovoltaic power generation system, determine DNI and component cleanliness as key factors affecting photovoltaic power generation. Because the data of component cleanliness cannot be directly obtained, the difference between component temperature and ambient temperature is used as a representative quantity of component cleanliness. DNI, component cleanliness, photovoltaic power generation power of the previous day and predicted photovoltaic power generation power are linearly related, and a linear regression algorithm is used to build a model for prediction.

5. The photovoltaic output prediction method considering haze / sand dust factors according to claim 1 is characterized in that: Step S4, using the echo state network ESN, set at time t, the input is x ( t ), the input node is m , the reserve pool status is y ( t ), the number of nodes is n , the output is z ( t ), the number of nodes is o , the state of each layer is expressed as: , T is the matrix transposition symbol. The state equations of each layer in the current stage are shown above, and the state equations of the next stage are obtained as follows: , in the formula f and f out They are the activation functions of the model state and output equations respectively. The default activation function is the tanh function. W in is the connection weight from the input layer to the reserve pool, W is the connection weight between the storage pools, W out is the connection weight between the reserve pool and the output layer, W back is the feedback output weight, x ( t +1) is the input of the next moment at time t, y ( t +1) is the state of the reserve pool at the next moment after time t, z ( t +1) is the output of the next moment after time t. From the state equation of the next stage of the reserve pool, it can be seen that the current state at each moment is related to the states of multiple moments, and is respectively related to the state response at the current moment and the state response at the previous stage. At the same time, during the update process, the output will be updated through the feedback matrix W back Read and adjust the information; Taking wind speed, visibility, PM2.5, PM10, and ambient temperature as input, DNI and component cleanliness as output, and the above data in the haze / dust day category in historical data as the training data set, the echo state network model, namely the haze / dust impact model 1, is obtained.

6. The photovoltaic output prediction method considering haze / sand dust factors according to claim 1 is characterized in that: Step S5, establish the following function: , , q is the input feature, h For output, θ is the characteristic parameter, ε is the error, θ T q (i) is the predicted value, r (i) is the true value, and the loss function is: , i Indicates the quantity number involved in the calculation, Q Represents the input feature matrix, and the least squares method is used to solve the feature parameters θ : ; Taking DNI, component cleanliness, and photovoltaic power generation of the previous day as input features, the predicted daily photovoltaic power generation is output, and the above data of haze / sandstorm categories in historical data are used as training data sets to obtain photovoltaic power generation prediction model 2.

7. The photovoltaic output prediction method considering haze / sand dust factors according to claim 1, characterized in that: Step S6, using the k-fold cross-validation method, divide the data set into k subsets, each subset is used as a test set, and the rest is used as a training set each time, cross-validation is performed k times, and the average cross-validation performance index of k times is used as the final result. The mean absolute error and root mean square error are used to measure the prediction performance of the model: Mean absolute error: ; Root Mean Square Error: ; Where a is the number of prediction results, A is the total number of calculations of the Σ symbol formula; P means (a) , P pred (a) are the actual values ​​and the model predicted values.

Images