A method for predicting a cleaning cycle of a photovoltaic module

Abstract

The application discloses a kind of photovoltaic module cleaning cycle prediction methods, first establish the power generation loss rate prediction model of photovoltaic module to obtain power generation loss rate, then determine cleaning threshold, according to the sum of the change value of the power generation loss rate corresponding to all collection time in cycle according to cleaning threshold, when exceed the range of cleaning threshold, turn into cleaning decision process, make the decision of cleaning by combining artificial and intelligent prediction, according to weak consciousness echo state network constructs power generation loss rate prediction model, and according to the change of input data, select optimal network weight, compared with other static neural network, its fitting ability is stronger, prediction accuracy is higher, the data set of dynamic acquisition is constantly updated to prediction model, avoid the problem that the prediction model needs to be modified due to the change of data characteristics in offline prediction, improve the reliability of prediction model, reach the effect of online prediction.

Description

Technical Field

[0001] This invention relates to the field of big data technology, and in particular to a method for predicting the cleaning cycle of photovoltaic modules. Background Technology

[0002] Photovoltaic energy is a clean energy source with enormous potential, and photovoltaic modules are the core component for generating photovoltaic energy. However, dust accumulation on photovoltaic modules reduces light transmittance, thus affecting the lifespan of the modules and the power generation of the photovoltaic power plant. Current technologies for photovoltaic power plant cleaning decisions generally fall into two categories: one is for maintenance personnel to observe the dust accumulation on the photovoltaic modules on-site and make judgments based on experience, but experience alone is rarely sufficient for a scientific assessment; the other is to judge based on the photovoltaic power generation figures, making decisions by comparing the difference in power generation between dust-accumulated and non-dust-accumulated modules. However, the degree of dust accumulation and power generation dynamically change with regional climate, dust accumulation conditions, and other factors. Natural conditions such as wind, sunlight, and temperature constantly influence cleaning decisions, creating a dynamic process. The dynamic updating of the dataset and the data analysis algorithm affect the analysis results, making it impossible to perform photovoltaic module cleaning on a fixed cycle. High-frequency cleaning of photovoltaic power plants also increases the operation and maintenance costs of photovoltaic modules. Therefore, a smart and efficient method for predicting the cleaning cycle of photovoltaic modules is needed to achieve higher power generation returns with limited investment and reduce the operation and maintenance costs of photovoltaic modules. Summary of the Invention

[0003] To address the above issues and overcome the shortcomings of existing technologies, this invention aims to provide a method for predicting the cleaning cycle of photovoltaic modules. It utilizes the surface structure of a weak-consciousness echo state network prediction method to train and obtain optimal network weights, which are then stored in a deep consciousness space. The invention proposes the concept of data labels based on input data, using these labels to retrieve the optimal network weights. By continuously updating the prediction model with these data labels, the accuracy of the prediction model is improved. Finally, by comparing δ with a cleaning threshold, the transition to a cleaning decision is determined, thereby improving the cleaning efficiency of photovoltaic modules and saving cleaning costs.

[0004] The technical solution is a method for predicting the cleaning cycle of photovoltaic modules, which includes the following specific steps:

[0005] S1. Establish a predictive model for the power generation loss rate of photovoltaic modules. Use a meteorological monitoring station to collect weather data at different times during the actual power generation of the photovoltaic module under test. The weather data includes solar irradiance, ambient temperature and humidity, and wind speed. Then, use an IV curve analyzer to measure the actual power generation of the photovoltaic module under test, denoted as... j represents the sequence number of the corresponding acquisition time. A prediction model for the photovoltaic power system is constructed based on a weak-consciousness echo state network. Weather data is used as the input data for the prediction model, and the predicted power generation of clean photovoltaic modules is used as the output of the weak-consciousness echo state network. The prediction results, including the predicted power generation of clean photovoltaic panels under different input data at different times, are used as a comparison group. The predicted power generation is denoted as P′. j Then utilize the actual power generation P j With the predicted power generation P′ j The power generation loss rate η at the j-th data collection time is obtained. j The calculation formula is as follows:

[0006]

[0007] S2. Determine the cleaning threshold, collect the power generation of the photovoltaic module under natural dust accumulation conditions without cleaning, and obtain the cleaning threshold of the photovoltaic module to be tested through calculation and analysis.

[0008] S3. The decision to initiate cleaning is based on a comparison of the calculated power generation loss rate of the photovoltaic module under test during continuous data acquisition with the cleaning threshold. First, the change in power generation loss rate during continuous data acquisition is calculated and denoted as Δη. j(j+1) The calculation formula is as follows:

[0009] Δη j(j+1) =η j+1 -η j ,

[0010] The sum of the changes in power generation loss rate corresponding to all data collection moments within a cleaning cycle is denoted as δ. The formula for calculating δ is as follows:

[0011]

[0012] Where M is the number of times from the end of a cleaning cycle to the current collection time. When δ exceeds the set cleaning threshold, the process of making a cleaning decision begins.

[0013] S4. Construct a cleaning decision and calculate the difference between power generation revenue and operation and maintenance cost within the cleaning interval corresponding to the data collection time. When the difference is the largest, submit the data to manual decision-making. If the manual decision-making determines that the cleaning conditions are not met, return to S2 to redetermine the cleaning threshold. Then, make a second judgment on the photovoltaic module to be tested based on the redetermined cleaning threshold. Repeat the judgment process until the manual judgment is made to complete the cleaning cycle.

[0014] The structure of the weak consciousness echo state network of S1 includes a surface structure and a deep consciousness space. The surface structure predicts power generation, and the deep consciousness space stores the optimal network weights corresponding to data labels. The surface structure includes an input layer, a reservoir, and an output layer. The number of neurons in the input layer, reservoir, and output layer are denoted as K, N, and L, respectively. The state update equation is as follows:

[0015] x(n+1)=f(w in u(n+1)+Wx(n))+w fb y(n),

[0016] Where x(n) is the state variable, n = 1, 2, 3...T are discrete time points, T is the number of data points in the training sample data, f is the reservoir state activation function, and y = y(n) is the output vector. Let W represent the input weight matrix, W∈R N×N×C This represents the internal weight matrix of the reserve pool. Represents the output feedback matrix. Called the deep consciousness space, C in C, C fb The depth, referred to as the deep consciousness space, stores the optimal network weights and corresponding data labels obtained from network training, W. in W, W fb The output equation is as follows:

[0017] y(n)=g(w out [x(n);u(n)]),

[0018] Where g is the output activation function, x(n) is the state variable, u(n) is the input vector, y = y(n) is the output vector, d = d(n) is the desired output vector, e is the network contrast error, n = 1, 2, 3...T are discrete time points, T is the number of data points in the training sample data, [;] is the concatenation of vectors, W out ∈R L×(K+N) This indicates the output weight matrix.

[0019] In S2, the power generation is used to represent the degree of dust accumulation of the photovoltaic module. The power generation under natural dust accumulation conditions is collected, and the decrease value in different time intervals is calculated as time increases. The power generation influence curve is plotted, the decrease rate of power generation over time is calculated, the maximum value of the decrease rate of power generation is selected, and the maximum value of the power generation decrease efficiency is set as the cleaning threshold.

[0020] The data collection cycle begins after this cleaning and ends before the next cleaning. If an anomaly occurs during the data collection process, resulting in insufficient data, the missing values ​​in the dataset are corrected using multiple imputation methods. Each variable containing a missing value is predicted by default using other variables in the dataset. An interpolation function is established to predict the valid values ​​of the missing data. This process is iterated until all missing values ​​converge.

[0021] The prediction model established in S1 using a weak-consciousness echo state network uses the power generation dataset of clean photovoltaic modules as training data. The power generation dataset includes weather data of clean photovoltaic modules and power generation data collected by an IV curve instrument. The training data of the prediction model is a static dataset. The autocorrelation and cross-correlation coefficients of the input data are used as data labels for the corresponding network weights of the weak-consciousness echo state network. The optimal network weights are obtained by training the training data. When the data labels of the input data change, the network is reconstructed according to the existing data labels, and the data collected each time is updated in the input data. When the input data has corresponding data labels, the weak-consciousness echo state network of the corresponding prediction model directly retrieves the optimal network weights according to the data labels, without the need for weight training.

[0022] Due to the adoption of the above technical solutions, the present invention has the following advantages compared with the prior art;

[0023] 1. This invention utilizes a weak-consciousness echo state network to predict the power generation of clean photovoltaic modules. The predicted power generation is then obtained by analyzing the actual power generation of the photovoltaic module under test against the predicted power generation at the corresponding acquisition time. The power generation loss rate is determined by comparing the changes in the power generation loss rate at all acquisition times within a cycle with the cleaning threshold. The transition to the cleaning decision stage is determined by comparing these changes with other neural networks. The weak-consciousness echo state network has a stronger fitting ability than other neural networks. When the input data changes, the corresponding data labels also change. The weak-consciousness echo state network dynamically adjusts the network weight structure according to the changes in the input data. Furthermore, for input data with the same data labels, the network directly retrieves the optimal network weights from the deep consciousness space, eliminating the need for retraining the prediction model and allowing for reconstruction of the prediction model in the shortest possible time, thereby improving the speed and efficiency of prediction.

[0024] 2. This invention sets up a control group by collecting and analyzing the power generation dataset of clean photovoltaic modules, which improves the richness of the comparative data. The power generation dataset of clean photovoltaic modules is used for network training, and then the predicted power generation of the photovoltaic modules under test in the clean state is predicted by weather data, which improves the accuracy of the control variables of the comparison group and improves the accuracy of the comparison results.

[0025] 3. A cleaning threshold is introduced, and power generation is used to judge the dust accumulation. This avoids the errors of judging the dust accumulation of photovoltaic panels by using existing image recognition and manual methods. The cumulative decrease and rate of decrease in power generation are used for judgment. When the cleaning threshold is passed and the cleaning decision process begins, if the manual decision does not meet the cleaning conditions, the cleaning threshold is re-evaluated, thus combining manual and intelligent prediction. Attached Figure Description

[0026] Figure 1 This is an overall flowchart of the present invention;

[0027] Figure 2 This is a structural diagram of the weak consciousness echo state network of the present invention;

[0028] Figure 3 This is a diagram of the surface structure of the weak consciousness echo state network of the present invention. Detailed Implementation

[0029] The foregoing and other technical contents, features and effects of the present invention are described in conjunction with the appendix below. Figures 1 to 3 The detailed description of the embodiments will make this clear. All structural details mentioned in the following embodiments are based on the accompanying drawings.

[0030] A method for predicting the cleaning cycle of photovoltaic modules includes the following specific steps:

[0031] S1. Establish a predictive model for the power generation loss rate of photovoltaic modules. Clean photovoltaic modules include clean photovoltaic panels. Select one or more clean photovoltaic panels as the experimental data collection group. Place these clean photovoltaic panels outdoors for the experiment. Use cleaning devices and other hardware facilities to clean the photovoltaic panels frequently to keep them in a relatively clean state. The photovoltaic modules to be tested include photovoltaic panels waiting to be judged as clean. Use a meteorological monitoring station to collect weather data corresponding to different times when the photovoltaic modules under test are actually generating power. The weather data includes solar irradiance, ambient temperature and humidity, and wind speed. Then use an IV curve instrument to measure the actual power generation of the photovoltaic modules under test, and record it as follows. j represents the sequence number of the corresponding acquisition time. A prediction model for the photovoltaic power system is constructed based on a weak-consciousness echo state network. Echo state networks, proposed in 2001, use a reservoir to replace the hidden layers of a traditional recurrent neural network (RNN), representing a novel type of RNN. It consists of an input layer, a reservoir, and an output layer. The weak-consciousness echo state network improves upon the echo state network by introducing the concept of "labels" into the reservoir. The structure of the weak-consciousness echo state network comprises two layers: a surface structure and a deep consciousness space. The surface structure represents the power generation prediction model, while the deep consciousness space stores the optimal network weights corresponding to the labels. By processing the input data, "data labels" are obtained for storage and retrieval. The network is then trained to generate labels. The optimal network weights corresponding to the labels are stored in the deep consciousness space. When the input data changes, the network can think autonomously, automatically search the consciousness space based on the defined labels, extract and use the optimal network weights, and reconstruct the network in the shortest possible time. In the surface structure, a prediction model is obtained by calculating the Spearman correlation coefficient between photovoltaic power generation data and corresponding weather data using an unsupervised learning algorithm of a restricted Boltzmann machine. Weather data is used as the input data for the prediction model, and the predicted power generation of clean photovoltaic modules is used as the output of the weak consciousness echo state network. The prediction model is trained using the power generation dataset of clean photovoltaic modules. The prediction results of the predicted power generation of clean photovoltaic panels under different input data at different times are used as a comparison group. The predicted power generation is denoted as P′. j Then utilize the actual power generation P j With the predicted power generation P' j The power generation loss rate η at the j-th data collection time is obtained. j The calculation formula is as follows:

[0032]

[0033] S2. Determine the cleaning threshold. Since the amount of dust accumulated on the photovoltaic panels of photovoltaic modules cannot be obtained by detection, we use power generation to represent the dust accumulation situation for judgment. We collect the power generation of the photovoltaic modules under natural dust accumulation conditions without cleaning, and calculate and analyze to obtain the cleaning threshold of the photovoltaic modules to be tested.

[0034] S3. The decision to initiate cleaning is based on a comparison of the calculated power generation loss rate of the photovoltaic module under test during continuous data acquisition with the cleaning threshold. First, the change in power generation loss rate during continuous data acquisition is calculated and denoted as Δη. j(j+1) The calculation formula is as follows:

[0035] Δη j(j+1) =η j+1 -η j ,

[0036] Within a cleaning cycle, as the time of data collection for the photovoltaic module under test progresses, the power generation loss rate of the photovoltaic module under test gradually increases over time, Δη j(j+1) Δη represents the magnitude of the change in the power generation loss rate. j(j+1) The smaller the value, the smaller the change in the power generation loss rate between two consecutive data collection times, Δη. j(j+1) The larger the value, the greater the change in the power generation loss rate between two consecutive measurement times. The sum of the changes in the power generation loss rate corresponding to all measurement times collected within a cleaning cycle is denoted as δ. The formula for calculating δ is as follows:

[0037]

[0038] Where M is the number of times from the end of a cleaning cycle to the current collection time. When δ exceeds the set cleaning threshold, the process of making a cleaning decision begins.

[0039] S4. Construct a cleaning decision and calculate the difference between power generation revenue and operation and maintenance cost within the cleaning interval corresponding to the data collection time. When the difference is the largest, submit the data to manual decision-making. If the manual decision-making determines that the cleaning conditions are not met, return to S2 to redetermine the cleaning threshold. Then, make a second judgment on the photovoltaic module to be tested based on the redetermined cleaning threshold. Repeat the judgment process until the manual judgment is made to complete the cleaning cycle.

[0040] The structure of the weak-consciousness echo state network in S1 includes a surface structure and a deep consciousness space. Weather data serves as the network input, and predicted power generation is the network output. Data labels are network labels. The surface structure predicts power generation, while the deep consciousness space stores the optimal network weights corresponding to the data labels. The surface structure includes an input layer, a storage pool, and an output layer. The surface structure obtains the optimal network weights through training data and input data, and stores them in the deep consciousness space. The optimal network weights are retrieved by searching. The number of neurons in the input layer, storage pool, and output layer are denoted as K, N, and L, respectively. The state update equation is as follows:

[0041] x(n+1)=f(w in u(n+1)+Wx(n))+w fb y(n),

[0042] Where x(n) is the state variable, n = 1, 2, 3...T are discrete time points, T is the number of data points in the training sample data, f is the reservoir state activation function, and y = y(n) is the output vector. Let W represent the input weight matrix, W∈R N×N×C This represents the internal weight matrix of the reserve pool. This represents the output feedback matrix. Called the deep consciousness space, C in C, C fb The depth, referred to as the deep consciousness space, stores the optimal network weights and corresponding data labels obtained from network training, W. in W, W fb In this process, when data labels exist for the input data, the network directly retrieves the optimal network weights stored in the deep consciousness space, saving the steps of training the prediction model. The output equation is as follows:

[0043] y(n)=g(w out [x(n);u(n)]),

[0044] Where g is the output activation function, x(n) is the state variable, u(n) is the input vector, y = y(n) is the output vector, d = d(n) is the desired output vector, e is the network contrast error, n = 1, 2, 3...T are discrete time points, T is the number of data points in the training sample data, [;] is the concatenation of vectors, W out ∈R L×(K+N) This indicates the output weight matrix.

[0045] In S2, the power generation is used to represent the degree of dust accumulation in the photovoltaic module. The power generation under natural dust accumulation conditions without cleaning the panel is collected, and the decrease value in different time intervals is calculated as time increases. The power generation influence curve is plotted, the decrease rate of power generation over time is calculated, the maximum value of the decrease rate of power generation is selected, and the maximum value of the power generation decrease efficiency is set as the cleaning threshold.

[0046] The data collection cycle begins after this cleaning and ends before the next cleaning. If an anomaly occurs during the data collection process, resulting in insufficient data, the missing values ​​in the dataset are corrected using multiple imputation methods. Each variable containing a missing value is predicted by default using other variables in the dataset. An interpolation function is established to predict the valid values ​​of the missing data. This process is iterated until all missing values ​​converge.

[0047] The prediction model established in S1 using a weak-consciousness echo state network uses a power generation dataset of clean photovoltaic modules as training data. The power generation dataset includes weather data of clean photovoltaic modules and power generation data collected by an IV curve instrument. The training data of the prediction model is a static dataset. The autocorrelation and cross-correlation coefficients of the input data are used as data labels for the corresponding network weights of the weak-consciousness echo state network. The optimal network weights are obtained by training the training data. When the data labels of the input data change, the network is reconstructed according to the existing data labels, and the data collected each time is updated in the input data. When the input data has a corresponding data label, the weak-consciousness echo state network of the corresponding prediction model directly retrieves the optimal network weights according to the data label, without the need for weight training. The prediction model is continuously updated through the data model, thereby improving the accuracy of the prediction and eliminating the training process.

[0048] In practical use, this invention mainly includes four steps: First, a prediction model for the power generation loss rate of photovoltaic modules is established to obtain the power generation loss rate. Second, a cleaning threshold is determined. Then, the sum of the changes in the power generation loss rate of the collected data points is judged based on the cleaning threshold. When δ exceeds the range of the cleaning threshold, the process enters the cleaning decision process. During the cleaning decision process, if the manual decision is not to clean, the photovoltaic module to be tested is judged again based on the newly determined cleaning threshold. The judgment process is repeated until the manual decision is made to clean, thus completing one cleaning cycle. The prediction model is continuously updated based on the dynamically collected dataset. For input data with the same data label, the network directly retrieves the optimal network weights stored in the deep consciousness space. By continuously updating the data labels and prediction model, the accuracy and timeliness of the prediction model are improved. The sum of the changes in the power generation loss rate is compared with the cleaning threshold to determine whether to enter the cleaning decision process. The weak consciousness echo state network constructs the power generation loss rate prediction model and selects the optimal network weights based on the changes in the input data. Compared with other static neural networks, it has stronger fitting ability and higher prediction accuracy. As data is collected under different seasons and weather conditions, the prediction model is continuously updated based on the collected data, avoiding the problem of needing to modify the prediction model due to changes in data characteristics in offline prediction. This improves the reliability of the prediction model, achieves the effect of online prediction, and reduces the operation and maintenance costs of photovoltaic modules.

[0049] The above description is a further detailed explanation of the present invention in conjunction with specific embodiments, and it should not be considered that the specific implementation of the present invention is limited to this. For those skilled in the art to which the present invention pertains and related fields, any extensions, operation methods, and data substitutions made based on the technical solution concept of the present invention should fall within the protection scope of the present invention.

Claims

1. A method for predicting the cleaning cycle of photovoltaic modules, characterized in that, The specific steps include the following: S1. Establish a predictive model for the power generation loss rate of photovoltaic modules. Use a meteorological monitoring station to collect weather data at different times during the actual power generation of the photovoltaic module under test. The weather data includes solar irradiance, ambient temperature and humidity, and wind speed. Then, use an IV curve analyzer to measure the actual power generation of the photovoltaic module under test, denoted as... , To correspond to the sequence number of the data collection time, a prediction model for the photovoltaic power system was constructed based on a weak-consciousness echo state network. Weather data was used as the input data for the prediction model, and the predicted power generation of clean photovoltaic modules was used as the output of the weak-consciousness echo state network. The prediction results, including the predicted power generation of clean photovoltaic panels under different input data at different times, were used as a comparison group. The predicted power generation was denoted as... Then utilize the actual power generation capacity With predicted power generation Get the first Power generation loss rate at each data collection time The calculation formula is as follows: ； S2. Determine the cleaning threshold, collect the power generation of the photovoltaic module under natural dust accumulation conditions without cleaning, and obtain the cleaning threshold of the photovoltaic module to be tested through calculation and analysis. S3. The decision to initiate cleaning is based on a comparison of the calculated power generation loss rate of the photovoltaic module under test during continuous data acquisition with the cleaning threshold. First, the change in power generation loss rate during continuous data acquisition is calculated and denoted as... The calculation formula is as follows: - ， The sum of the changes in power generation loss rate at all collected times within a cleaning cycle is denoted as... , The calculation formula is as follows: ， Where M is the number of times from the end of one cleaning cycle to the current data collection time, when If the set cleaning threshold is exceeded, the process will proceed to the cleaning decision-making stage. S4. Construct a cleaning decision and calculate the difference between power generation revenue and operation and maintenance cost within the cleaning interval corresponding to the data collection time. When the difference is the largest, submit the data to manual decision-making. If the manual decision-making determines that the cleaning conditions are not met, return to S2 to redetermine the cleaning threshold. Then, make a second judgment on the photovoltaic module to be tested based on the redetermined cleaning threshold. Repeat the judgment process until the manual judgment is made to complete a cleaning cycle. The structure of the weak consciousness echo state network in S1 includes a surface structure and a deep consciousness space. The surface structure predicts power generation, while the deep consciousness space stores the optimal network weights corresponding to data labels. The surface structure contains an input layer, a reservoir, and an output layer. The number of neurons in the input layer, reservoir, and output layer are denoted as follows: , and The state update equation is as follows: ， in, For state variables, For discrete time points, The number of data points in the training sample data. The activation function for the reserve pool state. For the output vector, This represents the input weight matrix. This represents the internal weight matrix of the reserve pool. This represents the output feedback matrix. , , This is called the deep consciousness space. , , The depth, referred to as the deep consciousness space, stores the optimal network weights obtained from network training, along with their corresponding data labels. , , The output equation is as follows: ， To output the activation function, For the input vector, For the desired output vector, For network contrast error, For the concatenation of vectors, This indicates the output weight matrix.

2. The method for predicting the cleaning cycle of photovoltaic modules according to claim 1, characterized in that, In S2, the power generation is used to represent the degree of dust accumulation of the photovoltaic module. The power generation under natural dust accumulation conditions is collected, and the decrease value in different time intervals is calculated as time increases. The power generation influence curve is plotted, the decrease rate of power generation over time is calculated, the maximum value of the decrease rate of power generation is selected, and the maximum value of the power generation decrease efficiency is set as the cleaning threshold.

3. The method for predicting the cleaning cycle of photovoltaic modules according to claim 1, characterized in that, The data collection cycle begins after this cleaning and ends before the next cleaning. If an anomaly occurs during the data collection process, resulting in insufficient data, the missing values ​​in the dataset are corrected using multiple imputation methods. Each variable containing a missing value is predicted by default using other variables in the dataset. An interpolation function is established to predict the valid values ​​of the missing data. This process is iterated until all missing values ​​converge.

4. The method for predicting the cleaning cycle of photovoltaic modules according to claim 1, characterized in that, The prediction model established in S1 using a weak-consciousness echo state network uses the power generation dataset of clean photovoltaic modules as training data. The power generation dataset includes weather data of clean photovoltaic modules and power generation data collected by an IV curve instrument. The training data of the prediction model is a static dataset. The autocorrelation and cross-correlation coefficients of the input data are used as data labels for the corresponding network weights of the weak-consciousness echo state network. The optimal network weights are obtained by training the training data. When the data labels of the input data change, the network is reconstructed according to the existing data labels, and the data collected each time is updated in the input data. When the input data has corresponding data labels, the weak-consciousness echo state network of the corresponding prediction model directly retrieves the optimal network weights according to the data labels, without the need for weight training.

Images