A smart photovoltaic monitoring method and system based on deep learning

Abstract

The present invention discloses an intelligent photovoltaic monitoring method and system based on deep learning, including S1: obtaining the light intensity, air temperature and original photovoltaic power data in a unit time and preprocessing them to obtain the preprocessed light intensity, air temperature and photovoltaic power data; S2: dividing the photovoltaic array into regions, and using the local outlier factor method to perform rough detection of abnormal values, and calculating the output power mean of all normal regions; S3: constructing a dynamic threshold training network, combining the real-time light intensity, air temperature and the output power mean of the normal region, and training the dynamic threshold; S4: setting the input, output and optimization target of the dynamic threshold training network; S5: optimizing the parameters of the dynamic threshold training network based on an adaptive optimization algorithm. The present invention solves the problem that the traditional fixed threshold cannot be adjusted adaptively, ensures that the photovoltaic monitoring system can accurately judge the operating status of the photovoltaic panel under different weather conditions, and improves the robustness and reliability of the system.

Description

Technical Field

[0001] The present invention belongs to the field of solar photovoltaic technology, and in particular relates to an intelligent photovoltaic monitoring method and system based on deep learning. Background Art

[0002] Smart photovoltaic monitoring uses intelligent technology and algorithms to monitor and analyze the performance and status of photovoltaic arrays in real time to ensure their efficient operation and detect potential problems in a timely manner. Traditional photovoltaic monitoring methods usually set a fixed power generation as the alarm threshold based on experience, and alarm is achieved by comparing the actual power generation of the photovoltaic array with the set fixed threshold. This method cannot adapt to power generation under different weather and environmental conditions and may result in false alarms or missed alarms. Summary of the invention

[0003] In view of this, the present invention aims to provide an intelligent photovoltaic monitoring method and system based on deep learning, aiming to solve the problem that the alarm threshold of the traditional photovoltaic monitoring method is fixed and cannot be adaptively adjusted.

[0004] A deep learning-based intelligent photovoltaic monitoring method comprises the following steps:

[0005] S1: Obtain the light intensity, temperature and original photovoltaic power data per unit time and preprocess them to obtain the preprocessed light intensity, temperature and photovoltaic power data;

[0006] S2: Divide the photovoltaic array into regions, use the local outlier factor method to perform rough detection of outliers, and calculate the average output power of all normal regions;

[0007] S3: Build a dynamic threshold training network, combine the real-time light intensity, temperature and the average output power of the normal area to train the dynamic threshold;

[0008] S4: Set the input, output and optimization target of the dynamic threshold training network;

[0009] S5: Optimize the parameters of the dynamic threshold training network based on an adaptive optimization algorithm.

[0010] Furthermore, the step S1 obtains the light intensity, air temperature and original photovoltaic power data within a unit time and performs preprocessing to obtain the preprocessed light intensity, air temperature and photovoltaic power data, including:

[0011] The preprocessing includes missing value filling and data normalization. The calculation method of missing value filling is:

[0012]

[0013] Among them, X pq is the data point in the pth row and qth column, is the value after filling, is the mean of the qth column;

[0014] The calculation method for data normalization is:

[0015]

[0016] Among them, X p ′ q is the normalized data point, min(X q ) is the minimum value of the qth column data, max(X q ) is the maximum value of the data in the qth column;

[0017] The light intensity, temperature and raw photovoltaic power data per unit time are usually continuous in form, that is, they are collected in chronological order, reflecting the trend and pattern of light intensity, temperature and photovoltaic power changes over time; and can take any value within a certain range.

[0018] Furthermore, the step S2 divides the photovoltaic array into regions, and uses the local outlier factor method to perform rough detection of outliers, and calculates the output power mean of all normal regions, including:

[0019] S21: Divide all photovoltaic arrays into n photovoltaic array regions, each region containing m photovoltaic arrays;

[0020] Unlike the traditional method that directly uses the average power of a few photovoltaic panels as the reference threshold, the present invention divides the entire photovoltaic array into multiple areas, each of which covers multiple photovoltaic panels. The advantage of this area division is that even if a problem occurs in a certain area, other areas can still provide reliable data for generating thresholds; since the probability of problems occurring in all areas at the same time is extremely low, selecting photovoltaic panels in normal areas as the data source for the threshold can ensure the reliability of the threshold, thereby improving the robustness and reliability of the photovoltaic monitoring system and effectively solving the problem of failure of the reference threshold;

[0021]

[0022] Where LOF(i) is the LOF value, i is the current data point, M(i) is the set of adjacent data points, || is the potential of the set, density(j) is the density of adjacent data point j, and density(i) is the density of the current data point;

[0023] Compared with complex anomaly detection algorithms, the calculation process of the LOF method is relatively simple, mainly involving the calculation of the distance and density between data points, without the need for complex mathematical operations or iterative optimization. Therefore, the LOF method has a faster calculation speed and is suitable for real-time monitoring systems. In the present invention, since the alarm threshold needs to be adaptively adjusted in real time with the power of each photovoltaic array, a rough abnormal value detection method with high calculation efficiency and the ability to be quickly updated in a real-time monitoring system is required. The simple and efficient calculation process of the LOF method enables data to be processed in a short time and results to be obtained quickly, thereby realizing timely monitoring of the photovoltaic array status.

[0024] S23: Eliminate the photovoltaic array area containing abnormal values, and calculate the output power average of all normal areas respectively.

[0025] Furthermore, the step S3 constructs a dynamic threshold training network, combines the real-time light intensity, air temperature and the average output power of the normal area, and trains the dynamic threshold, including:

[0026] S31: Construct a time series weight extractor to extract the importance of light intensity, temperature and the average output power of the normal area at different time steps. The calculation method is:

[0027] Attention(h t )=softmax(W att tanh(W h ·h t +b h ))

[0028] Among them, Attention(h t ) is the attention weight, h t is the hidden state sequence based on the recurrent neural network, softmax() is the softmax function, W att is the attention weight parameter, tanh() is the tanh function, W h is the parameter matrix for mapping the hidden state sequence, b h is the bias vector for mapping the hidden state sequence;

[0029] Time series weight extraction is achieved by calculating attention weights. Such a design can dynamically focus on key information at different time steps to more accurately determine the alarm threshold. In the photovoltaic array monitoring scenario, weather conditions will change significantly during the day and night, and on sunny and cloudy days. If the information at all time steps is simply regarded as equally important, false alarms will increase when the lighting conditions are good, and problems will not be discovered in time when the lighting conditions are poor.

[0030] S32: Combine the attention weight to correct the time series, and the calculation method is:

[0031]

[0032] Among them, WeightedInput(x t ,Attention(h t )) is the time series after weight correction, x t is the original input data, N is the number of dimensions of the time series data, f is the time series data index, x t,f is the original input data with index, Attention(h t ) f is the indexed attention weight;

[0033] S33: Construct a time series feature extractor to extract the time series variation characteristics of light intensity, temperature and output power average in normal area. The calculation method is:

[0034] Feature(x t )=GRU(WeightedInput(x t ,Attention(h t )))

[0035] Among them, Feature(x t ) is the time series change feature, and GRU() is the GRU network operation; the GRU network can effectively capture the time dependency in the data, thereby more accurately reflecting the time series changes of the light intensity, temperature and the average output power in the normal area;

[0036] Compared with the traditional method of directly inputting raw time series data, the present invention uses the time series sequence corrected by attention weight as input, which can reduce the interference of less important time series information, improve the accuracy and effectiveness of feature extraction, better reflect the time series change characteristics of the photovoltaic monitoring system, and provide a more reliable basis for the accurate calculation of dynamic thresholds;

[0037] S34: Integrate the time series variation characteristics of light intensity, temperature and output power in normal area, and output the alarm threshold. The calculation method is:

[0038] Threshold=Linear(Concatenate([Feature(x t )1,Feature(x t )2,Feature(x t )3]))

[0039] Among them, Threshold is the alarm threshold, Linear() is the neural network linear layer operation, Concatenate() is the concatenation operation, Feature(xt )1,Feature(x t )2,Feature(x t )3 are the time series variation characteristics of light intensity, temperature and output power average in normal area respectively;

[0040] By splicing various features and processing them through a linear layer, the final alarm threshold is obtained, which can fully take into account the temporal variation characteristics of light intensity, temperature and the average output power in the normal area, so as to more accurately determine the dynamic alarm threshold to adapt to the working status of the photovoltaic system under different time periods and environmental conditions.

[0041] Furthermore, the step S4 sets the input, output and optimization target of the dynamic threshold training network, including:

[0042] S41: setting the input of the dynamic threshold training network to be the light intensity, the temperature and the average output power of the normal area, and the output of the dynamic threshold training network to be the alarm threshold;

[0043] S42: Set the loss function of the temporal weight extractor, which is calculated as follows:

[0044]

[0045] Among them, L att is the loss function of the temporal weight extractor, T is the sequence length, The attention weights predicted by the model;

[0046] S43: Set the loss function of the temporal feature extractor, which is calculated as follows:

[0047]

[0048] Among them, L gru is the loss function of the temporal feature extractor, Threshold(x t ) is the actual alarm threshold, is the alarm threshold predicted by the model, ∥∥ 2 is the square of the norm;

[0049] S44: The final loss function of the dynamic threshold training network is calculated as:

[0050] L total =L att +L gru

[0051] Among them, L total The final loss function for the dynamic threshold training network;

[0052] The loss function L of the temporal weight extractor attIt is used to measure the difference between the attention weight predicted by the model and the actual attention weight. By calculating the cross entropy loss, it evaluates the accuracy of the model's prediction of the importance of time series data at different time steps. The loss function L of the time series feature extractor is gru The accuracy of the model in extracting time series features is evaluated by calculating the square of the Euclidean distance between the alarm threshold predicted by the model and the actual alarm threshold; the accuracy of the model in extracting time series features is evaluated by minimizing the total loss function L total , the timing weight extractor and timing feature extractor can be optimized simultaneously to improve the performance and stability of the monitoring system.

[0053] Furthermore, the step S5 optimizes the parameters of the dynamic threshold training network based on an adaptive optimization algorithm, including:

[0054] L total =L att +L gru

[0055] Among them, L total Final loss function for dynamic threshold training network.

[0056] Furthermore, the step S5 optimizes the parameters of the dynamic threshold training network based on an adaptive optimization algorithm, including:

[0057] S51: Optimize the parameters of the dynamic threshold training network based on the RMSProp optimizer. The calculation method is:

[0058]

[0059]

[0060] Among them, r c is the cumulative squared gradient of the cth iteration, g c is the gradient of the cth iteration, θ c is the parameter of the dynamic threshold training network for the cth iteration, α is the global learning rate, ρ is the decay coefficient, and ∈ is a constant;

[0061] S52: Outputting alarm thresholds based on dynamic threshold training network;

[0062] S53: Early warning of the photovoltaic array in combination with the alarm threshold;

[0063] The RMSProp optimization algorithm adapts each parameter by adjusting the learning rate α to make the training process more stable; The normalization of the cumulative sum square root helps prevent the gradient from disappearing or exploding during training; therefore, compared with the standard gradient descent method, RMSProp can accelerate the convergence of the neural network;

[0064] Among them, S52 and S53 are in a progressive relationship; step S52 is to output an alarm threshold based on the dynamic threshold training network, and step S53 is to use this alarm threshold to warn the photovoltaic array.

[0065] The present invention also provides an intelligent photovoltaic monitoring system based on deep learning, comprising:

[0066] Data acquisition and preprocessing module: acquires the light intensity, temperature and original photovoltaic power data within a unit time and performs preprocessing to obtain the preprocessed light intensity, temperature and photovoltaic power data;

[0067] Photovoltaic array rough detection module: divides the photovoltaic array into regions, uses the local outlier factor method to perform rough detection of outliers, and calculates the average output power of all normal regions;

[0068] Dynamic threshold training module: build a dynamic threshold training network, combine the real-time light intensity, temperature and the average output power of the normal area to train the dynamic threshold;

[0069] Network training module: set the input, output and optimization target of the dynamic threshold training network;

[0070] Network optimization module: Optimize the parameters of the dynamic threshold training network based on an adaptive optimization algorithm to realize the above-mentioned intelligent photovoltaic monitoring method based on deep learning.

[0071] Compared with the prior art, the advantages of the present invention are:

[0072] (1) In terms of the overall solution, the present invention divides the photovoltaic array into multiple areas and dynamically calculates the threshold value based on the power average of the normal area, thereby solving the problem that the traditional fixed threshold value cannot be adaptively adjusted, ensuring that the photovoltaic monitoring system can accurately judge the operating status of the photovoltaic panel under different weather conditions, thereby improving the robustness and reliability of the system.

[0073] (2) In terms of algorithm improvement, the present invention proposes a dynamic threshold training network. By assigning different weights to the data at each time step, the information at different time steps in the time series data is fully utilized, and the timing characteristics of light intensity, air temperature and photovoltaic array output power are more flexibly captured, thereby improving the monitoring system's ability to accurately judge the operating status of photovoltaic panels.

[0074] (3) In terms of data utilization, the present invention divides the photovoltaic array into multiple areas for monitoring, and combines the local outlier factor method to perform rough detection of outliers; the regionalized data utilization method effectively avoids the occurrence of overall monitoring failure due to problems in a single benchmark area, and ensures that the monitoring system can still work normally when problems occur in a certain area. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] Figure 1 A schematic diagram of the process flow of the deep learning-based intelligent photovoltaic monitoring method provided by the present invention. DETAILED DESCRIPTION

[0076] The present invention is further described below in conjunction with the accompanying drawings, but the present invention is not limited in any way. Any changes or substitutions made based on the teachings of the present invention belong to the protection scope of the present invention.

[0077] Embodiment 1: A smart photovoltaic monitoring method based on deep learning, such as Figure 1 As shown, the following steps are included:

[0078] S1: Obtain the light intensity, temperature and original photovoltaic power data per unit time and preprocess them to obtain the preprocessed light intensity, temperature and photovoltaic power data;

[0079] The preprocessing includes missing value filling and data normalization. The calculation method of missing value filling is:

[0080]

[0081] Among them, X pq is the data point in the pth row and qth column, is the value after filling, is the mean of the qth column;

[0082] The calculation method for data normalization is:

[0083]

[0084] Among them, X p ′ q is the normalized data point, min(X q ) is the minimum value of the qth column data, max(X q ) is the maximum value of the data in the qth column;

[0085] The light intensity, temperature and raw photovoltaic power data per unit time are usually continuous in form, that is, they are collected in chronological order, reflecting the trend and pattern of light intensity, temperature and photovoltaic power changes over time; and can take any value within a certain range.

[0086] S2: Divide the photovoltaic array into regions, use the local outlier factor method to perform rough detection of outliers, and calculate the average output power of all normal regions;

[0087] S21: Divide all photovoltaic arrays into n photovoltaic array regions, each region containing m photovoltaic arrays;

[0088] Unlike the traditional method that directly uses the average power of a few photovoltaic panels as the reference threshold, the present invention divides the entire photovoltaic array into multiple areas, each of which covers multiple photovoltaic panels. The advantage of this area division is that even if a problem occurs in a certain area, other areas can still provide reliable data for generating thresholds; since the probability of problems occurring in all areas at the same time is extremely low, selecting photovoltaic panels in normal areas as the data source for the threshold can ensure the reliability of the threshold, thereby improving the robustness and reliability of the photovoltaic monitoring system and effectively solving the problem of failure of the reference threshold;

[0089] S22: For each photovoltaic array area, the local outlier factor method is used to perform rough outlier detection, and the calculation method is:

[0090]

[0091] Where LOF(i) is the LOF value, i is the current data point, M(i) is the set of adjacent data points, || is the potential of the set, density(j) is the density of adjacent data point j, and density(i) is the density of the current data point;

[0092] Compared with complex anomaly detection algorithms, the calculation process of the LOF method is relatively simple, mainly involving the calculation of the distance and density between data points, without the need for complex mathematical operations or iterative optimization. Therefore, the LOF method has a faster calculation speed and is suitable for real-time monitoring systems. In the present invention, since the alarm threshold needs to be adaptively adjusted in real time with the power of each photovoltaic array, a rough abnormal value detection method with high calculation efficiency and the ability to be quickly updated in a real-time monitoring system is required. The simple and efficient calculation process of the LOF method enables data to be processed in a short time and results to be obtained quickly, thereby realizing timely monitoring of the photovoltaic array status.

[0093] S23: Eliminate the photovoltaic array area containing abnormal values, and calculate the output power average of all normal areas, for example:

[0094] For example, the entire photovoltaic array is divided into 4 areas, each area contains 5 photovoltaic panels, for one of the areas:

[0095] Photovoltaic panel 1: 100W

[0096] Photovoltaic panel 2: 95W

[0097] Photovoltaic panel 3: 98W

[0098] Photovoltaic panel 4: 102W

[0099] Photovoltaic panel 5: 50W (abnormal value)

[0100] For PV panel 5 (outlier):

[0101] Assuming that density is directly proportional to power, the LOF value is:

[0102]

[0103]

[0104] LOF(5)=1.975, which is much larger than 1, indicating that the output power of photovoltaic panel 5 is an abnormal value compared with other photovoltaic panels; the area where photovoltaic panel No. 5 is located is eliminated, and then the output power averages of the other three normal areas are calculated.

[0105] S3: Build a dynamic threshold training network, combine the real-time light intensity, temperature and the average output power of the normal area to train the dynamic threshold;

[0106] S31: Construct a time series weight extractor to extract the importance of light intensity, temperature and the average output power of the normal area at different time steps. The calculation method is:

[0107] Attention(h t )=softmax(W att tanh(W h ·h t +b h ))

[0108] Among them, Attention(h t ) is the attention weight, h t is the hidden state sequence based on the recurrent neural network, softmax() is the softmax function, W att is the attention weight parameter, tanh() is the tanh function, W h is the parameter matrix for mapping the hidden state sequence, b h is the bias vector for mapping the hidden state sequence;

[0109] Time series weight extraction is achieved by calculating attention weights. Such a design can dynamically focus on key information at different time steps to more accurately determine the alarm threshold. In the photovoltaic array monitoring scenario, weather conditions will change significantly during the day and night, and on sunny and cloudy days. If the information at all time steps is simply regarded as equally important, false alarms will increase when the lighting conditions are good, and problems will not be discovered in time when the lighting conditions are poor.

[0110] S32: Combine the attention weight to correct the time series, and the calculation method is:

[0111]

[0112] Among them, WeightedInput(x t ,Attention(h t )) is the time series after weight correction, x t is the original input data, N is the number of dimensions of the time series data, f is the time series data index, x t,f is the original input data with index, Attention(h t ) f is the indexed attention weight;

[0113] S33: Construct a time series feature extractor to extract the time series variation characteristics of light intensity, temperature and output power average in normal area. The calculation method is:

[0114] Feature(x t )=GRU(WeightedInput(x t ,Attention(h t )))

[0115] Among them, Feature(x t ) is the time series change feature, and GRU() is the GRU network operation; the GRU network can effectively capture the time dependency in the data, thereby more accurately reflecting the time series changes of the light intensity, temperature and the average output power in the normal area;

[0116] Compared with the traditional method of directly inputting raw time series data, the present invention uses the time series sequence corrected by attention weight as input, which can reduce the interference of less important time series information, improve the accuracy and effectiveness of feature extraction, better reflect the time series change characteristics of the photovoltaic monitoring system, and provide a more reliable basis for the accurate calculation of dynamic thresholds;

[0117] S34: Integrate the time series variation characteristics of light intensity, temperature and output power in normal area, and output the alarm threshold. The calculation method is:

[0118] Threshold=Linear(Concatenate([Feature(x t )1,Feature(x t )2,Feature(x t )3]))

[0119] Among them, Threshold is the alarm threshold, Linear() is the neural network linear layer operation, Concatenate() is the concatenation operation, Feature(x t )1,Feature(x t )2,Feature(xt )3 are the time series variation characteristics of light intensity, temperature and output power average in normal area respectively;

[0120] By splicing various features and processing them through a linear layer, the final alarm threshold is obtained, which can fully take into account the time series variation characteristics of light intensity, temperature and the average output power in the normal area, so as to more accurately determine the dynamic alarm threshold to adapt to the working status of the photovoltaic system under different time periods and environmental conditions;

[0121] For example, suppose the photovoltaic power station is located in an area with changeable climate, where the light intensity and temperature vary greatly during the day, especially in the afternoon, when cloud changes may cause the light intensity to drop rapidly. Therefore, for the monitoring system of the photovoltaic power station, the data in the afternoon is more important. Assume the data is as follows:

[0122] Before noon (time step (t-2)): light intensity 500Lux, temperature 20℃, power 300W

[0123] Noon (time step (t-1)): light intensity 800Lux, temperature 30℃, power 500W

[0124] Afternoon (time step (t)): light intensity 600Lux, temperature 25℃, power 400W

[0125] The attention mechanism can learn the weights at each time step based on the characteristics of data changes. In this example, the weights are set to 0.2, 0.3, and 0.5.

[0126] Then the corrected data input to GRU is:

[0127] Before noon (time step (t-2)): [100,4,60]

[0128] Noon (time step (t-1)): [240,9,150]

[0129] Afternoon (time step (t)): [300, 12.5, 200]

[0130] The weighted data emphasizes the importance of the afternoon time period. This dynamic adjustment enables the monitoring system to adjust the threshold according to changes in real-time data, thereby more accurately reflecting the working status of the PV system at critical moments.

[0131] S4: Set the input, output and optimization target of the dynamic threshold training network;

[0132] S41: setting the input of the dynamic threshold training network to be the light intensity, the temperature and the average output power of the normal area, and the output of the dynamic threshold training network to be the alarm threshold;

[0133] S42: Set the loss function of the temporal weight extractor, which is calculated as follows:

[0134]

[0135] Among them, L att is the loss function of the temporal weight extractor, T is the sequence length, The attention weights predicted by the model;

[0136] S43: Set the loss function of the temporal feature extractor, which is calculated as follows:

[0137]

[0138] Among them, L gru is the loss function of the temporal feature extractor, Threshold(x t ) is the actual alarm threshold, is the alarm threshold predicted by the model, ∥∥ 2 is the square of the norm;

[0139] S44: The final loss function of the dynamic threshold training network is calculated as:

[0140] L total =L att +L gru

[0141] Among them, L total The final loss function for the dynamic threshold training network;

[0142] The loss function L of the temporal weight extractor att It is used to measure the difference between the attention weight predicted by the model and the actual attention weight. By calculating the cross entropy loss, it evaluates the accuracy of the model's prediction of the importance of time series data at different time steps. The loss function L of the time series feature extractor is gru The accuracy of the model in extracting time series features is evaluated by calculating the square of the Euclidean distance between the alarm threshold predicted by the model and the actual alarm threshold; the accuracy of the model in extracting time series features is evaluated by minimizing the total loss function L total , the timing weight extractor and timing feature extractor can be optimized simultaneously to improve the performance and stability of the monitoring system.

[0143] S5: Optimizing the parameters of the dynamic threshold training network based on an adaptive optimization algorithm;

[0144] S51: Optimize the parameters of the dynamic threshold training network based on the RMSProp optimizer. The calculation method is:

[0145]

[0146]

[0147] Among them, r c is the cumulative squared gradient of the cth iteration, g c is the gradient of the cth iteration, θ c is the parameter of the dynamic threshold training network for the cth iteration, α is the global learning rate, ρ is the decay coefficient, and ∈ is a constant;

[0148] S52: Outputting alarm thresholds based on dynamic threshold training network;

[0149] S53: Early warning of the photovoltaic array in combination with the alarm threshold;

[0150] The RMSProp optimization algorithm adapts each parameter by adjusting the learning rate α to make the training process more stable; The normalization of the cumulative sum square root helps prevent the gradient from disappearing or exploding during training; therefore, compared with the standard gradient descent method, RMSProp can accelerate the convergence of the neural network;

[0151] Among them, S52 and S53 are in a progressive relationship; step S52 is to output an alarm threshold based on the dynamic threshold training network, and step S53 is to use this alarm threshold to warn the photovoltaic array.

[0152] Embodiment 2: The present invention also discloses a deep learning-based intelligent photovoltaic monitoring system, which is used to implement the above-mentioned deep learning-based intelligent photovoltaic monitoring method, including:

[0153] Data acquisition and preprocessing module: acquires the light intensity, temperature and original photovoltaic power data within a unit time and performs preprocessing to obtain the preprocessed light intensity, temperature and photovoltaic power data;

[0154] Photovoltaic array rough detection module: divides the photovoltaic array into regions, uses the local outlier factor method to perform rough detection of outliers, and calculates the average output power of all normal regions;

[0155] Dynamic threshold training module: build a dynamic threshold training network, combine the real-time light intensity, temperature and the average output power of the normal area to train the dynamic threshold;

[0156] Network training module: set the input, output and optimization target of the dynamic threshold training network;

[0157] Network optimization module: Optimizes the parameters of the dynamic threshold training network based on an adaptive optimization algorithm.

[0158] It should be noted that the serial numbers of the above embodiments of the present invention are only for description and do not represent the advantages and disadvantages of the embodiments. And the terms "including", "comprising" or any other variants thereof in this article are intended to cover non-exclusive inclusion, so that a process, device, article or method including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, device, article or method. In the absence of further restrictions, an element defined by the sentence "including a ..." does not exclude the presence of other identical elements in the process, device, article or method including the element.

[0159] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes a number of instructions for a terminal device (which can be a mobile phone, a computer, a server, or a network device, etc.) to execute the methods described in each embodiment of the present invention.

[0160] The above are only preferred embodiments of the present invention, and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A deep learning-based intelligent photovoltaic monitoring method, comprising collecting light intensity and photovoltaic array output power data using a photovoltaic array composed of multiple photovoltaic panels connected mechanically and electrically, characterized in that: The following steps are involved: S1: Obtain the light intensity, temperature and original photovoltaic power data per unit time and preprocess them to obtain the preprocessed light intensity, temperature and photovoltaic power data; S2: Divide the photovoltaic array into regions, use the local outlier factor method to perform rough detection of outliers, and calculate the average output power of all normal regions; S21: Divide all photovoltaic arrays into n photovoltaic array regions, each region containing m photovoltaic arrays; S22: For each photovoltaic array area, the local outlier factor method is used to perform rough outlier detection; S23: Eliminate the photovoltaic array area containing abnormal values, and calculate the output power average of all normal areas respectively; S3: Build a dynamic threshold training network, combine the real-time light intensity, temperature and the average output power of the normal area to train the dynamic threshold; including the following steps: S31: Construct a time series weight extractor to extract the importance of light intensity, temperature and the average output power of the normal area at different time steps. The calculation method is: Attention(h t )=softmax(W att ·tanh(W h ·h t +b h )) Among them, Attention(h t ) is the attention weight, h t is the hidden state sequence based on the recurrent neural network, softmax() is the softmax function, W att is the attention weight parameter, tanh() is the tanh function, W h is the parameter matrix for mapping the hidden state sequence, b h is the bias vector for mapping the hidden state sequence; S32: Combine the attention weight to correct the time series, and the calculation method is: Among them, WeightedInput(x t ,Attention(h t )) is the time series after weight correction, x t is the original input data, N is the number of dimensions of the time series data, f is the time series data index, x t，f is the original input data with index, Attention(h t ) f is the indexed attention weight; S33: Construct a time series feature extractor to extract the time series variation characteristics of light intensity, temperature and output power average in normal area. The calculation method is: Feature(x t )=GRU(WeightedInput(x t ,Attention(h t ))) Among them, Feature(x t ) is the time series change feature, GRU() is the GRU network operation; S34: Integrate the time series variation characteristics of light intensity, temperature and output power in normal area, and output the alarm threshold. The calculation method is: Threshold=Linear(Concatenate([Feature(x t )1,Feature(x t )2,Feature(x t )3])) Among them, Threshold is the alarm threshold, Linear() is the neural network linear layer operation, Concatenate() is the concatenation operation, Feature(x t )1, Feature(x t )2, Feature(x t )3 are the time series variation characteristics of light intensity, temperature and output power average in normal area respectively; S4: Set the input, output and optimization target of the dynamic threshold training network; S5: Optimize the parameters of the dynamic threshold training network based on an adaptive optimization algorithm.

2. The deep learning-based intelligent photovoltaic monitoring method according to claim 1, characterized in that: The step S1 comprises the following steps: The preprocessing includes missing value filling and data normalization. The calculation method of missing value filling is: Among them, X pq is the data point in the pth row and qth column, is the value after filling, is the mean of the qth column; The calculation method for data normalization is: Among them, X′ pq is the normalized data point, min(X q ) is the minimum value of the qth column data, max(X q ) is the maximum value of the data in the qth column.

3. The deep learning-based intelligent photovoltaic monitoring method according to claim 2 is characterized in that: The step S2 comprises the following steps: S21: Divide all photovoltaic arrays into n photovoltaic array regions, each region containing m photovoltaic arrays; S22: For each photovoltaic array area, the local outlier factor method is used to perform rough outlier detection, and the calculation method is: Where LOF(i) is the LOF value, i is the current data point, M(i) is the set of adjacent data points, || is the potential of the set, density(j) is the density of adjacent data point j, and density(i) is the density of the current data point; S23: Eliminate the photovoltaic array area containing abnormal values, and calculate the output power average of all normal areas respectively.

4. The deep learning-based intelligent photovoltaic monitoring method according to claim 1, characterized in that: The step S4 comprises the following steps: S41: setting the input of the dynamic threshold training network to be the light intensity, the temperature and the average output power of the normal area, and the output of the dynamic threshold training network to be the alarm threshold; S42: Set the loss function of the temporal weight extractor, which is calculated as follows: Among them, L att is the loss function of the temporal weight extractor, T is the sequence length, The attention weights predicted by the model; S43: Set the loss function of the temporal feature extractor, which is calculated as follows: Among them, L gru is the loss function of the temporal feature extractor, Threshold(x t ) is the actual alarm threshold, is the alarm threshold predicted by the model, || || 2 is the square of the norm; S44: The final loss function of the dynamic threshold training network is calculated as: L total =L att +L gru Among them, L total Final loss function for dynamic threshold training network.

5. The deep learning-based intelligent photovoltaic monitoring method according to claim 4 is characterized in that: The step S5 comprises the following steps: S51: Optimize the parameters of the dynamic threshold training network based on the RMSProp optimizer. The calculation method is: Among them, r c is the cumulative squared gradient of the cth iteration, g c is the gradient of the cth iteration, θ c is the parameter of the dynamic threshold training network for the cth iteration, α is the global learning rate, ρ is the decay coefficient, and ∈ is a constant; S52: Outputting alarm thresholds based on dynamic threshold training network; S53: Early warning of the photovoltaic array is performed in combination with the alarm threshold.

6. A deep learning-based intelligent photovoltaic monitoring system, characterized in that: include: Data acquisition and preprocessing module: acquires the light intensity, temperature and original photovoltaic power data within a unit time and performs preprocessing to obtain the preprocessed light intensity, temperature and photovoltaic power data; Photovoltaic array rough detection module: divides the photovoltaic array into regions, uses the local outlier factor method to perform rough detection of outliers, and calculates the average output power of all normal regions; Dynamic threshold training module: build a dynamic threshold training network, combine the real-time light intensity, temperature and the average output power of the normal area to train the dynamic threshold; Network training module: set the input, output and optimization target of the dynamic threshold training network; Network optimization module: optimizes the parameters of the dynamic threshold training network based on the adaptive optimization algorithm; To realize the deep learning-based intelligent photovoltaic monitoring method as described in any one of claims 1-5.

Images