A power station operation and maintenance state monitoring method based on time series data multi-scale features

Abstract

This invention discloses a method for monitoring the operation and maintenance status of power plants based on multi-scale features of time-series data, comprising the following steps: acquiring current monitoring data and historical monitoring data of all photovoltaic (PV) devices; calculating environmental correction factors and equipment correction factors based on the current and historical monitoring data of all PV devices; normalizing the current monitoring data of all PV devices based on the environmental and equipment correction factors; inputting the normalized monitoring data as time-series data into multiple time-series convolutional units to obtain a feature sequence with multi-scale features; passing the multi-scale feature sequence through two layers of long short-term memory (LSTM) units to obtain a sequence feature vector; passing the sequence feature vector through two layers of fully connected networks to obtain an output vector; and using the Softmax function to calculate the fault probability of the PV devices. This invention improves the operation and maintenance capabilities of PV power plants and increases the accuracy of diagnosing power plant faults.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic power plant operation and maintenance technology, and in particular to a method for monitoring the operation and maintenance status of power plants based on multi-scale characteristics of time-series data. Background Technology

[0002] Photovoltaic power generation plays a vital role in my country's new energy development strategy. Distributed photovoltaic (PV) power stations are a type of PV infrastructure, referring to PV facilities built near user sites where users consume the generated electricity themselves and feed surplus power into the grid. Distributed PV power stations can alleviate localized power shortages and offer advantages such as being noise-free and pollution-free.

[0003] Current power plant operation and maintenance monitoring algorithms can be broadly categorized into two types: those based on equipment electrical characteristics and those based on machine learning. Algorithms based on equipment electrical characteristics rely on specific and accurate electrical data for mathematical modeling. Obtaining this data requires complex hardware support and manual maintenance, making widespread application in actual production difficult. Machine learning-based diagnostic algorithms can extract features of faulty equipment from different types of input data and can tolerate noise in the input data. However, existing operation and maintenance monitoring algorithm research relies on the assumption that there are no differences between equipment and weather conditions on different dates. This assumption only applies to centralized photovoltaic power plants, while in distributed photovoltaic power plants, there are significant differences in equipment location, placement angle, and commissioning time. Furthermore, existing machine learning algorithms often only utilize four data points: AC current, AC voltage, DC current, and DC voltage, while actual monitoring data often contains more than ten items, containing a wealth of information that can be mined and utilized.

[0004] Therefore, existing technologies have high requirements for the actual application environment or low accuracy in diagnosing equipment faults, making it difficult to meet the actual needs of operation and maintenance monitoring of distributed photovoltaic power stations. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a power plant operation and maintenance status monitoring method based on multi-scale features of time-series data. This method reduces the environmental requirements for practical applications and improves the diagnostic accuracy of equipment faults.

[0006] This invention is achieved through the following technical solution:

[0007] A method for monitoring the operation and maintenance status of power plants based on multi-scale features of time-series data includes the following steps:

[0008] Obtain current and historical monitoring data for all photovoltaic devices;

[0009] Environmental correction factors are calculated based on current monitoring data of all photovoltaic devices, and equipment correction factors are calculated based on historical monitoring data of all photovoltaic devices.

[0010] Based on the aforementioned environmental correction factors and equipment correction factors, the current monitoring data of all photovoltaic devices are standardized to obtain standardized monitoring data;

[0011] The normalized monitoring data is input as time-series data into multiple time-series convolutional units to obtain a feature sequence with multi-scale features;

[0012] The feature sequence of the multi-scale feature is passed through two layers of long short-term memory units to obtain a sequence feature vector;

[0013] The sequence feature vector is passed through two fully connected layers to obtain the output vector, and the output vector is processed by the Softmax function to obtain the failure probability of the photovoltaic equipment.

[0014] Furthermore, the historical monitoring data for all photovoltaic devices includes monitoring data from at least the past year.

[0015] Furthermore, the steps described above—calculating environmental correction factors based on current monitoring data of all photovoltaic devices and calculating equipment correction factors based on historical monitoring data of all photovoltaic devices—specifically include the following steps:

[0016] Calculate the average value of a certain electrical data at various times of the day, that is:

[0017]

[0018] Among them, w i,j,k Let j be a certain electrical data point monitored by the j-th group of equipment on day i at time k. This represents the average value of this data at various times throughout the day, where K is the number of times data is recorded each day.

[0019] Calculate the environmental correction factor for day i. Right now:

[0020]

[0021] Where J represents the number of devices;

[0022] Calculate the equipment correction factor for the j-th group of equipment. Right now:

[0023]

[0024] Where I represents the number of dates with recorded data.

[0025] Further, the step of standardizing the current monitoring data of all photovoltaic devices based on the environmental correction factor and the equipment correction factor to obtain standardized monitoring data specifically includes:

[0026] Environmental correction factors on day i and the correction factor of the jth group of equipment The monitoring data w of the j-th group of equipment on day i at time k i,j,k Convert to relative values Right now:

[0027]

[0028] Among them, relative values This refers to the standardized monitoring data.

[0029] Furthermore, the step of inputting the normalized monitoring data as time-series data into multiple time-series convolutional units to obtain a feature sequence with multi-scale features specifically includes the following steps:

[0030] Temporal data is input into three parallel temporal convolutional units with different kernel sizes.

[0031] Neighborhood features at different scales were extracted respectively;

[0032] The neighborhood features at different scales are spliced ​​together to form a feature sequence with multi-scale features.

[0033] Furthermore, the sizes of the three temporal convolutional units are 1×3, 1×5, and 1×7, respectively.

[0034] Furthermore, the temporal convolutional unit includes a convolutional layer, a batch normalization layer, and a ReLU activation function.

[0035] Furthermore, the two-layer long short-term memory unit includes a first long short-term memory unit and a second long short-term memory unit.

[0036] Furthermore, an attention unit is disposed between the first long short-term memory unit and the second long short-term memory unit.

[0037] Furthermore, the step of obtaining a sequence feature vector by passing the feature sequence of the multi-scale features through two layers of long short-term memory units specifically includes the following steps:

[0038] The feature sequence of the multi-scale features is input into the first long short-term memory unit to form the output iterative sequence;

[0039] An enhanced intermediate representation sequence is calculated using the attention mechanism of the attention unit and based on the iterative sequence;

[0040] The enhanced intermediate representation sequence is input into the second long short-term memory unit, and the output sequence feature vector is generated.

[0041] Compared with the prior art, the advantages of this invention are:

[0042] 1. This invention overcomes the problem of existing research techniques relying on the assumption that there are no differences between equipment and no differences in environmental conditions across different dates. Starting from the monitoring data itself, it calculates environmental correction factors and equipment correction factors, and standardizes the monitoring data through horizontal and vertical comparisons, correcting the interference caused by equipment differences and environmental differences in equipment monitoring.

[0043] 2. This invention utilizes convolutional units of different scales to propose multi-scale temporal features, smoothing noise in the data while reducing information loss. Compared to existing methods, it is more suitable for processing distributed power station monitoring data with large fluctuations and high noise levels.

[0044] 3. Compared with existing technologies, this invention more fully mines the sequence information in the monitoring data. It uses two layers of long short-term memory units and strengthens the connection between the two layers through an attention mechanism, thereby improving the diagnostic accuracy of faulty equipment. Attached Figure Description

[0045] Figure 1 This is a flowchart of a power plant operation and maintenance status monitoring method based on multi-scale features of time-series data according to the present invention;

[0046] Figure 2 This is a schematic diagram of the temporal convolution structure of the present invention. Detailed Implementation

[0047] The following detailed, non-limiting description of the invention's technical solutions, in conjunction with preferred embodiments and accompanying drawings, is provided. In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," "counterclockwise," "axial," "radial," and "circumferential," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the invention, and should not be construed as limiting the invention.

[0048] like Figure 1 As shown, an embodiment of the present invention provides a method for monitoring the operation and maintenance status of a power plant based on multi-scale features of time-series data, comprising the following steps:

[0049] S1: Obtain the current monitoring data and historical monitoring data of all photovoltaic devices respectively.

[0050] Specifically, the current and historical monitoring data of all photovoltaic equipment are obtained through the power station's monitoring system. The historical monitoring data of all photovoltaic equipment includes monitoring data from at least the past year.

[0051] The monitoring system of a distributed photovoltaic (PV) power station measures and records various electrical parameters of the PV power generation equipment at a certain frequency, including power generation, AC side voltage, AC side current, DC side voltage, and DC side current. To ensure the validity of the monitoring data, each parameter is recorded only once every 15 minutes between 6:00 AM and 8:00 PM. The electrical data for the day form their own data sequences. This method also requires electrical data recorded in the same manner over the past year.

[0052] S2: Calculate the environmental correction factor based on the current monitoring data of all photovoltaic devices, and calculate the equipment correction factor based on the historical monitoring data of all photovoltaic devices.

[0053] Step S2 specifically includes the following steps:

[0054] S20: Calculate the average value of a certain electrical data at various times of the day, i.e.:

[0055]

[0056] Among them, w i,j,k Let j be a certain electrical data point monitored by the j-th group of equipment on day i at time k. This represents the average value of this data at various times throughout the day, where K is the number of times data is recorded each day.

[0057] Taking electricity generation as an example, then w i,j,k This represents the power generation monitored by the j-th group of devices on day i at the k-th time of that day; This represents the power generation level of the j-th group of equipment on day i.

[0058] S21: Calculate the environmental correction factor for day i. Right now:

[0059]

[0060] Where J represents the number of devices;

[0061] S22: Calculate the equipment correction factor for the j-th group of equipment. Right now:

[0062]

[0063] Where I represents the number of dates with recorded data.

[0064] The above steps perform data standardization processing on each monitoring data point. Taking power generation as an example, the standardization process is performed using w... i,j,k This represents the power generation monitored by the j-th group of devices on day i at the k-th time of that day; The average power generation at various times throughout the day is used to measure the power generation level of the j-th group of equipment on day i; based on this, the environmental correction factor for day i is calculated. and the equipment correction factor of the jth group of equipment

[0065] S3: Standardize the current monitoring data of all photovoltaic devices based on environmental correction factors and equipment correction factors to obtain standardized monitoring data.

[0066] Specifically, this includes environmental correction factors on day i. and the correction factor of the jth group of equipment The monitoring data w of the j-th group of equipment on day i at time k i,j,k Convert to relative values Right now:

[0067]

[0068] Among them, relative values This refers to the standardized monitoring data.

[0069] The above steps overcome the problem of existing research techniques relying on the assumption that there are no differences between equipment and no differences in environmental conditions across different dates. Starting from the monitoring data itself, environmental correction factors and equipment correction factors are calculated, and the monitoring data is standardized through horizontal and vertical comparisons to correct for interference caused by differences in equipment and environment in equipment monitoring.

[0070] S4: The normalized monitoring data is input as time-series data into multiple time-series convolutional units to obtain a feature sequence with multi-scale features.

[0071] Step S4 specifically includes the following steps:

[0072] S40: Temporal data is input into three parallel temporal convolutional units with different kernel sizes;

[0073] The three temporal convolutional units have sizes of 1×3, 1×5, and 1×7, and each temporal convolutional unit includes a convolutional layer, a batch normalization layer, and a ReLU activation function.

[0074] S41: Extract neighborhood features at different scales respectively;

[0075] S42: Connect neighborhood features at different scales to form a feature sequence with multi-scale features.

[0076] The specific operations for steps S40 to S42 are as follows:

[0077] To address the characteristics of large fluctuations and high noise in monitoring data from distributed photovoltaic power plants, multiple parallel temporal convolutions are used to extract neighborhood features at different scales. This smooths the input data while reducing the loss of input information. The structure of a single temporal convolution unit is as follows: Figure 2 As shown, it consists of convolutional layers, batch normalization layers, and a ReLU activation function, and its operation can be represented as:

[0078]

[0079] s = BN(e)

[0080] h = ReLU(s)

[0081] Where x is the input of the temporal convolutional unit and h is the output of the temporal convolutional unit.

[0082] Multiple parallel temporal convolutional units can be described as:

[0083] h1 = C 1×3 (x)

[0084] h2=C 1×5 (x)

[0085] h3 = C 1×7 (x)

[0086] y = cat(h1, h2, h3)

[0087] Among them, C 1×3 C 1×5 C 1×7 These represent temporal convolutional units with kernel sizes of 1×3, 1×5, and 1×7, respectively. `cat(.)` represents feature concatenation, which concatenates h1, h2, and h3 into a single vector, where the dimension of the resulting vector is the sum of the dimensions of the three vectors. `y` represents the output feature sequence with multi-scale features.

[0088] The above steps utilize convolutional units of different scales to extract multi-scale temporal features, smoothing noise in the data while reducing information loss. Compared to existing methods, this approach is more suitable for processing distributed power plant monitoring data with large fluctuations and high noise levels.

[0089] S5: The feature sequence of multi-scale features is passed through two layers of long short-term memory units to obtain the sequence feature vector.

[0090] Specifically, the two-layer long short-term memory unit includes a first long short-term memory unit and a second long short-term memory unit, with an attention unit located between the first long short-term memory unit and the second long short-term memory unit.

[0091] Step S50 specifically includes the following steps:

[0092] S50: The feature sequence of multi-scale features is input into the first long short-term memory unit to form the output iterative sequence;

[0093] S51: The enhanced intermediate representation sequence is calculated through the attention mechanism of the attention unit and based on the iteration sequence;

[0094] Specifically, the attention mechanism unit obtains the input of the second long short-term memory unit by weighted summing of the output of the first long short-term memory unit in each iteration.

[0095] S52: Input the enhanced intermediate representation sequence into the second long short-term memory unit and output the sequence feature vector.

[0096] The specific operations for steps S50-S51 are as follows:

[0097] The feature sequence of the multi-scale features obtained in step S4 is input into the first long short-term memory unit to form the output iterative sequence, calculated as follows:

[0098] m1,m2,…,m k =LSTM1(x1,x2,…x k )

[0099] Where LSTM1 is the first Long Short-Term Memory unit, x1, x2, ... x k Let m1, m2, ... m be the feature sequence of the multi-scale features obtained in step S4. k This represents the iterative sequence obtained from the LSTM1 iterative computation.

[0100] The attention mechanism is achieved through attention units and based on the iterative sequence m1, m2, ... m k Calculate the enhanced intermediate representation sequence c1, c2, ... c k The specific calculation method is as follows:

[0101]

[0102] a ij =attention(c j-1 ,m j )

[0103] Here, attention(·) represents the attention mechanism network, implemented by a single fully connected layer, with an input dimension of c. j-1 and h j The sum of the dimensions of a, b, and c, results in an output dimension of 1. ij In the i-th iteration, m represents... j The weights. When calculating C1, let C0 be the zero vector.

[0104] The enhanced intermediate representation sequence c1, c2, ... c k It is sent to the second long short-term memory unit, namely:

[0105] y1,y2,…,y k =LSTM2(c1,c2,…c k )

[0106] Where LSTM2 is the first Long Short-Term Memory unit, y1, y2, ..., y k The output sequence of the second long short-term memory unit, y k That is, the last item in the output sequence is the final sequence feature vector, which serves as the input feature for subsequent fault diagnosis.

[0107] Compared to existing technologies, this invention more fully mines the sequence information in the monitoring data. It utilizes two layers of long short-term memory units and strengthens the connection between the two layers through an attention mechanism, thereby improving the diagnostic accuracy of faulty equipment.

[0108] S6: The sequence feature vector is passed through two fully connected layers to obtain the output vector. The output vector is then processed by the Softmax function to obtain the failure probability of the photovoltaic equipment.

[0109] Specifically, fault diagnosis is achieved through a two-layer fully connected network, which converts the sequence feature vector y obtained in step S5 into a single layer. k As input, an output vector is obtained, with a dimension of (P+1), where P is the number of photovoltaic equipment fault categories. The output vector is processed by the Softmax function, converting its values ​​into probabilities, where the first dimension represents the probability of the equipment operating normally, and the other P dimensions represent the probability of each type of fault.

[0110] This invention can be applied to real-world photovoltaic power plant operation and maintenance monitoring scenarios, automatically analyzing photovoltaic power plant monitoring data, detecting faulty photovoltaic equipment, and achieving a high detection accuracy rate.

[0111] This invention utilizes a power plant operation and maintenance status monitoring method based on multi-scale features of monitoring data to overcome external environmental problems such as large differences in equipment and large changes in environmental conditions within distributed photovoltaic power plants, thereby improving the operation and maintenance capabilities of photovoltaic power plants and enhancing the accuracy of photovoltaic equipment fault diagnosis.

[0112] The parts of this invention not described in detail are well-known in the art.

[0113] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A method for monitoring the operation and maintenance status of a power plant based on multi-scale features of time-series data, characterized in that, Includes the following steps: Obtain current and historical monitoring data for all photovoltaic devices; Environmental correction factors are calculated based on current monitoring data of all photovoltaic devices, and equipment correction factors are calculated based on historical monitoring data of all photovoltaic devices. The steps include: calculating the environmental correction factor based on the current monitoring data of all photovoltaic devices, and calculating the device correction factor based on the historical monitoring data of all photovoltaic devices. Specifically, this includes the following steps: Calculate the average value of a certain electrical data at various times throughout the day, i.e.: in, Let j be a certain electrical data point monitored by the j-th group of equipment on day i at time k. This represents the average value of this data at various times throughout the day, where K is the number of times data is recorded each day. Calculate the environmental correction factor for day i. ,Right now: Where J represents the number of devices; Calculate the equipment correction factor for the j-th group of equipment. ,Right now: Where I represents the number of dates with recorded data; Based on the aforementioned environmental correction factors and equipment correction factors, the current monitoring data of all photovoltaic devices are standardized to obtain standardized monitoring data; The normalized monitoring data is input as time-series data into multiple time-series convolutional units to obtain a feature sequence with multi-scale features; The feature sequence of the multi-scale feature is passed through two layers of long short-term memory units to obtain a sequence feature vector; The sequence feature vector is passed through two fully connected layers to obtain the output vector, and the output vector is processed by the Softmax function to obtain the failure probability of the photovoltaic equipment.

2. The power plant operation and maintenance status monitoring method according to claim 1, characterized in that, The historical monitoring data for all photovoltaic devices includes monitoring data from at least the past year.

3. The power plant operation and maintenance status monitoring method according to claim 1, characterized in that, The step involves standardizing the current monitoring data of all photovoltaic devices based on the environmental correction factor and the equipment correction factor to obtain standardized monitoring data, specifically including: Environmental correction factors on day i and the correction factor of the jth group of equipment The monitoring data of the j-th group of devices on day i at time k Convert to relative values ,Right now: Among them, relative values This refers to the standardized monitoring data.

4. The power plant operation and maintenance status monitoring method according to claim 1, characterized in that, The step involves inputting the normalized monitoring data as time-series data into multiple time-series convolutional units to obtain a feature sequence with multi-scale features. Specifically, this includes the following steps: Temporal data is input into three parallel temporal convolutional units with different kernel sizes. Neighborhood features at different scales were extracted respectively; The neighborhood features at different scales are spliced ​​together to form a feature sequence with multi-scale features.

5. The power plant operation and maintenance status monitoring method according to claim 4, characterized in that, The sizes of the three temporal convolutional units are 1×3, 1×5, and 1×7, respectively.

6. The power plant operation and maintenance status monitoring method according to claim 4, characterized in that, The temporal convolutional unit includes a convolutional layer, a batch normalization layer, and a ReLU activation function.

7. The power plant operation and maintenance status monitoring method according to claim 1, characterized in that, The two-layer long short-term memory unit includes a first long short-term memory unit and a second long short-term memory unit.

8. The power plant operation and maintenance status monitoring method according to claim 7, characterized in that, An attention unit is disposed between the first long short-term memory unit and the second long short-term memory unit.

9. The power plant operation and maintenance status monitoring method according to claim 8, characterized in that, The step described above involves obtaining a sequence feature vector by passing the feature sequence of the multi-scale features through two layers of long short-term memory units, specifically including the following steps: The feature sequence of the multi-scale features is input into the first long short-term memory unit to form the output iterative sequence; An enhanced intermediate representation sequence is calculated using the attention mechanism of the attention unit and based on the iterative sequence; The enhanced intermediate representation sequence is input into the second long short-term memory unit, and the output sequence feature vector is generated.

Images