Operation economy evaluation method and system based on big data analysis

Abstract

The invention discloses an operation economy evaluation method and system based on big data analysis, and relates to the technical field of electric power engineering, and the method comprises the steps: carrying out the data smoothing of a photovoltaic data set through a time sequence convolution auto-encoder, and obtaining a processed photovoltaic data set; performing photovoltaic data dimension analysis on the processed photovoltaic data set by using an LSTM-based operation cost analysis model to obtain photovoltaic operation and maintenance cost, and performing generating capacity prediction on the processed photovoltaic data set by using a photovoltaic generating capacity prediction model to obtain generating capacity of power station equipment. Inputting the generating capacity of the power station equipment into the alarm evaluation model based on Transform for data analysis to obtain the generating capacity retrieved in unit time, and finally calculating the operation and maintenance income based on the generating capacity retrieved in unit time and the photovoltaic operation and maintenance cost to evaluate the overall work income. The economical efficiency of the whole operation and maintenance scheme is evaluated by comprehensively evaluating the operation and maintenance cost and the generating capacity recovery condition, the optimal operation and maintenance scheme is helped to be formulated, and the operation and maintenance efficiency and economical efficiency of the power station are improved.

Description

Technical Field

[0001] The present application relates to the field of electric power engineering technology, and more specifically, to a method and system for evaluating the economic performance of an operation based on big data analysis. Background Art

[0002] As the world's attention to renewable energy and environmental protection issues continues to increase, the photovoltaic industry, as an important part of the new energy field, has a very broad long-term development prospect. However, as the grid-connected electricity price of power stations decreases year by year, the power generation income of power stations has gradually decreased, which makes reducing the operating cost of power stations and increasing operating income the key factors affecting the economic benefits of power stations. At present, most photovoltaic power stations are built in relatively remote areas, which leads to high operation and maintenance costs. The operation and maintenance of photovoltaic power stations mainly focus on two aspects: one is to ensure the safe operation of the power station, and the other is to repair equipment failures and other failures that affect power generation. Ensuring the safe operation of the power station mainly involves lightning protection failures, fire warnings, short circuits and other failures that affect the safe production of the power station. Once such failures are discovered, they must be handled immediately. Equipment failures and other failures that affect power generation are more complicated, including failures reported by the photovoltaic equipment itself, equipment failures discovered through data analysis, failures discovered during inspections such as shading and snow accumulation, and other failures that affect power generation. There are many types of these failures, and how to prioritize them and where to start repairing them has become a major challenge for power station operation and maintenance.

[0003] In addition, there are many problems in the management of operating costs of power plant operation and maintenance. At present, the operating costs of many power plants cannot be effectively calculated and counted, resulting in uncontrollable operation and maintenance costs and difficulty in effectively reducing them. At the same time, most power plant operation and maintenance operations do not fully consider their economic feasibility. The main purpose of operation and maintenance is to repair alarms rather than to evaluate the actual benefits brought by operation and maintenance operations. This practice not only leads to inefficient and chaotic operation and maintenance operations, but may also cause the cost of operation input to exceed the benefits brought by the operation, thereby affecting the overall economic benefits of the power plant.

[0004] Therefore, how to effectively manage operating costs, reduce operation and maintenance costs, how to evaluate the economic feasibility of operation and maintenance operations, improve operating efficiency, and increase operating profits have become key issues that need to be urgently addressed in the current operation and maintenance of photovoltaic power stations. Summary of the invention

[0005] The purpose of this application is to provide an operation economic evaluation method and system based on big data analysis in order to overcome the existing technical defects. By evaluating the economic feasibility of the overall operation and maintenance plan, it helps power plant management to formulate the optimal operation and maintenance plan, change the chaotic and aimless operation and maintenance mode of the power plant, and make the operation and maintenance of the power plant efficient and simple.

[0006] The purpose of this application is achieved through the following technical solutions: In the first aspect, the present application proposes a method for evaluating the economic performance of an operation based on big data analysis, the method comprising: The photovoltaic data set is smoothed using a time series convolutional autoencoder to obtain a processed photovoltaic data set, which includes historical operation and maintenance records, alarm information, equipment status, environmental factors, labor costs, transportation costs, spare parts, and tool costs. The LSTM-based operation cost analysis model is used to analyze the photovoltaic data dimension of the processed photovoltaic data set to obtain the photovoltaic operation and maintenance cost; Use the photovoltaic power generation prediction model to predict the power generation of the processed photovoltaic data set to obtain the power generation of the power station equipment; Input the power generation of power station equipment into the Transformer-based alarm evaluation model for data analysis to obtain the power generation saved per unit time; The operation and maintenance benefits are calculated based on the power generation recovered per unit time and the photovoltaic operation and maintenance costs to evaluate the overall benefits of the operation.

[0007] In one possible implementation, a temporal convolutional autoencoder includes an autoencoder and a temporal convolutional network. The autoencoder learns the optimal representation of input features of a photovoltaic data set by minimizing reconstruction errors through self-supervised learning. The temporal convolutional network is used to utilize the parallel processing capabilities of convolutional neural networks for data processing, uses dilated convolutions to capture long-term dependencies, and learns temporal features by performing convolution operations on a one-dimensional time axis.

[0008] In one possible implementation, the autoencoder includes a first encoder, a first decoder, and a flatten layer. The first encoder is used to compress high-dimensional input data into a low-dimensional latent representation, the first decoder is used to restore the low-dimensional latent representation into high-dimensional output data, and the flatten layer is used to flatten the output of the first encoder and then input it into the first decoder.

[0009] In one possible implementation, the first encoder includes four residual layers, each residual layer consisting of two 1-DCNN layers; The first decoder consists of two dense layers using the LeakyReLU activation function to introduce a negative slope.

[0010] In a possible implementation, the data smoothing process includes: Select the window size and slide the window point by point along the time axis. A new window is generated every time it moves one time unit. Take the average of the initial w data points, repeat w times, and generate supplementary data at the starting position; Average the last w-1 data points and repeat w-1 times to create supplementary data for the end positions.

[0011] In a possible implementation, the LSTM-based activity-based cost analysis model includes a memory unit and a gate structure, wherein the memory unit is used to transmit and store information throughout the sequence and maintain long-term memory; The gate structure includes input gate, forget gate, and output gate; Input gate, used to add information to the memory cell, controls the inflow of information through the sigmoid activation function; The forget gate is used to delete information from the memory unit and use the sigmoid activation function to control the discard of information; The output gate is used to determine the information output in the memory unit and uses the sigmoid activation function to control the outflow of information.

[0012] In a possible implementation, the photovoltaic power generation prediction model includes a second encoder and a second decoder; The second encoder includes a multi-head sparse self-attention mechanism and a distillation algorithm. The calculation formula of the second encoder is: ,in , , is the matrix obtained by performing different linear transformations on the input vector, is obtained by probabilistic sparse method. The resulting matrix, are the dimensions of the query vector and key vector, is the normalized activation function; The second decoder consists of two identical multi-head attention layers. The calculation formula of the second decoder is: ,in Indicates the second decoder An input sequence, Mark is The start marker of a sequence, is a placeholder. It is a splicing operation.

[0013] In a possible implementation, the Transformer-based alarm assessment model includes a third encoder, a third decoder, and a self-attention mechanism; The third encoder includes multiple identical stacked layers, each of which contains a multi-head self-attention mechanism and a fully connected feedforward network. The outputs of the multi-head self-attention mechanism and the fully connected feedforward network are processed by residual connection and layer normalization. The third decoder also includes multiple identical stacked layers, each of which includes a masked multi-head self-attention mechanism, a multi-head self-attention mechanism, and a fully connected feedforward network; The self-attention mechanism uses an attention function to map query and key-value pairs to an output, which is calculated as a weighted sum of values, with the weights calculated by a compatibility function between the query and the corresponding key.

[0014] In a possible implementation, operation and maintenance benefit = recovery benefit per unit time * evaluation time - photovoltaic operation and maintenance cost, wherein the recovery benefit per unit time is obtained based on the power generation recovered per unit time.

[0015] In a second aspect, the present application also proposes an operation economics evaluation system based on big data analysis, the system comprising: A data smoothing module is used to perform data smoothing on the photovoltaic data set using a time series convolutional autoencoder to obtain a processed photovoltaic data set; The first analysis module is used to use the LSTM-based operation cost analysis model to perform photovoltaic data dimension analysis on the processed photovoltaic data set to obtain the photovoltaic operation and maintenance cost; A power generation prediction module is used to use a photovoltaic power generation prediction model to predict the power generation of the processed photovoltaic data set to obtain the power generation of the power station equipment; The second module is used to input the power generation of power station equipment into the Transformer-based alarm evaluation model for data analysis to obtain the power generation saved per unit time; The calculation module is used to calculate the operation and maintenance benefits based on the power generation saved per unit time and the photovoltaic operation and maintenance costs to evaluate the overall operation benefits. The above main scheme of the present application and its further options can be freely combined to form multiple schemes, all of which are schemes that can be adopted and claimed for protection in the present application; and in the present application, (non-conflicting options) options and other options can also be freely combined. After understanding the scheme of the present application, those skilled in the art can understand that there are many combinations based on the existing technology and common knowledge, all of which are technical schemes to be protected by the present application, and they are not exhaustively listed here.

[0016] The present application discloses a method and system for evaluating the economic efficiency of an operation based on big data analysis, which relates to the field of electric power engineering technology. First, a time series convolutional autoencoder is used to perform data smoothing on a photovoltaic data set to obtain a processed photovoltaic data set. Second, an operation cost analysis model based on LSTM is used to perform photovoltaic data dimension analysis on the processed photovoltaic data set to obtain photovoltaic operation and maintenance costs. A photovoltaic power generation prediction model is used to predict the power generation of the processed photovoltaic data set to obtain the power generation of power station equipment. The power generation of power station equipment is input into a Transformer-based alarm evaluation model for data analysis to obtain the power generation recovered per unit time. Finally, the operation and maintenance benefits are calculated based on the power generation recovered per unit time and the photovoltaic operation and maintenance cost to evaluate the overall operation benefits. By comprehensively evaluating the operation and maintenance costs and the power generation recovery, the economic efficiency of the overall operation and maintenance plan is evaluated, which helps to formulate the optimal operation and maintenance plan and improves the efficiency and economy of power station operation and maintenance. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the embodiments will be briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work.

[0018] Figure 1 A flow chart of a method for evaluating the economic performance of an operation based on big data analysis proposed in an embodiment of the present application is shown.

[0019] Figure 2 A schematic diagram of the structure of the autoencoder proposed in an embodiment of the present application is shown; Figure 3 The structure diagram of the temporal convolutional network proposed in the embodiment of the present application is shown; Figure 4 The structure diagram of the temporal convolutional autoencoder proposed in the embodiment of the present application is shown; Figure 5 A schematic diagram of the structure of the sliding window proposed in the embodiment of the present application is shown; Figure 6 A diagram showing the smoothing effect of TCN-AE data proposed in an embodiment of the present application is shown; Figure 7 The Transformer-based model structure proposed in the embodiment of the present application is shown; Figure 8 The structure diagram of the Transformer model proposed in the embodiment of the present application is shown; Fig. 9 A flow chart of photovoltaic prediction, alarm evaluation and operation cost analysis proposed in an embodiment of the present application is shown. DETAILED DESCRIPTION

[0020] The following describes the embodiments of the present application through specific examples, and those skilled in the art can easily understand other advantages and effects of the present application from the contents disclosed in this specification. The present application can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present application. It should be noted that the following embodiments and features in the embodiments can be combined with each other without conflict.

[0021] Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making any creative work shall fall within the scope of protection of this application.

[0022] In the existing technology, the operating costs of power plant operation and maintenance are not effectively managed, and the operating costs cannot be calculated and counted, resulting in uncontrollable operation and maintenance costs and an inability to effectively reduce operation and maintenance costs. In addition, most power plant operation and maintenance operations do not take their economic feasibility into consideration, and the operation and maintenance is only for the purpose of repairing alarms. Whether the operation and maintenance operations bring benefits is not evaluated, resulting in inefficient and chaotic operation and maintenance operations and the inability to maximize benefits. There may even be a situation where the operation input costs are greater than the operation benefits. Therefore, how to manage and reduce operating costs, how to evaluate the economic feasibility of operation and maintenance operations, improve operation efficiency, and increase operation benefits are issues that urgently need to be solved in the current power plant operation and maintenance.

[0023] Therefore, in order to solve the above-mentioned technical problems, the embodiment of the present application proposes an operation economic evaluation method and system based on big data analysis, which helps power plant management to formulate the optimal operation and maintenance plan through economic evaluation of the overall operation and maintenance plan, changes the chaotic and purposeless operation and maintenance mode of the power plant, and makes the operation and maintenance of the power plant efficient and simple. It is described in detail below.

[0024] Please refer to Figure 1 , Figure 1 A flowchart of a method for evaluating the economic performance of an operation based on big data analysis proposed in an embodiment of the present application is shown. The method comprises the following steps: Step S100: Use a time series convolutional autoencoder to perform data smoothing on the photovoltaic data set to obtain a processed photovoltaic data set.

[0025] The processed PV data set includes historical operation and maintenance records, alarm information, equipment status, environmental factors, labor costs, transportation costs, spare parts, and tool costs.

[0026] Before smoothing the photovoltaic data, it is necessary to first collect relevant data from power stations that have been in operation for more than one year. The historical operation and maintenance records are the daily maintenance and inspection of the power station, including maintenance time, maintenance content, maintenance personnel and other information. Cost data includes the construction and operation costs of the power station, such as equipment procurement costs, labor costs, maintenance costs, etc. Equipment information is the model, specification, installation date, manufacturer and other information of various equipment in the power station. Alarm information is various alarms generated during the operation of the power station, including alarm type, alarm time, handling measures, etc. The power generation is the power generation data of the power station on a daily, monthly and annual basis, including total power generation, peak power generation, etc.

[0027] In the data preprocessing stage, the collected raw data must first be cleaned to ensure data quality and consistency. Data cleaning includes removing duplicate values, processing missing values, and detecting and processing outliers. The specific method of removing duplicate values ​​is to check whether there are exactly the same records in the data set and delete them. The specific method of processing missing values ​​is to choose to fill (using the mean, median, interpolation) or delete records with missing values. The specific method of detecting and processing outliers is to identify outliers (extreme values, error values) in the data and take corresponding processing measures, such as elimination, correction or replacement.

[0028] Although the quality of the data after preliminary cleaning has improved, there may still be noise and fluctuations. In order to further improve the quality of the data, a time series convolutional autoencoder (TCN-AE) can be used for data smoothing. First, a TCN-AE model is constructed, which includes an encoder and a decoder. The TCN-AE model is trained with preprocessed data, and the model parameters are optimized by minimizing the difference between the input data and the reconstructed data. The preprocessed photovoltaic data is input into the trained TCN-AE model, and the smoothed data is output.

[0029] The temporal convolutional autoencoder includes an autoencoder and a temporal convolutional network. The autoencoder learns the optimal representation of the input features of the photovoltaic dataset by minimizing the reconstruction error through self-supervised learning. The temporal convolutional network is used to utilize the parallel processing capabilities of convolutional neural networks for data processing, uses dilated convolutions to capture long-term dependencies, and learns temporal features by performing convolution operations on a one-dimensional time axis.

[0030] In an autoencoder, efficient encoding of data is learned through self-supervised learning. A compressed representation (encoding) is learned, and the original input is reconstructed from this compressed representation. By minimizing the reconstruction error (mean square error) between the input data and the reconstructed data, the autoencoder can learn the main features of the data.

[0031] The autoencoder includes a first encoder, a first decoder and a flatten layer. The first encoder is used to compress high-dimensional input data into a low-dimensional latent representation, the first decoder is used to restore the low-dimensional latent representation to high-dimensional output data, and the flatten layer is used to flatten the output of the first encoder and then input it into the first decoder.

[0032] Figure 2 A schematic diagram of the structure of the autoencoder proposed in an embodiment of the present application is shown, including an encoder-decoder neural network having symmetrically arranged feedforward neural network (FFNN) layers and a dimension based on a central intermediate layer. The same input data can be used for self-supervised learning of the input layer and the output layer, making full use of the inherent functions of deep neural networks to enhance nonlinear feature extraction. The autoencoder has better performance in data smoothing processing.

[0033] The input of the first encoder is one-dimensional feature data, which uses the data itself as a label and learns the optimal representation of the input feature distribution by minimizing the reconstruction error, thereby retaining the most important part of the input. The first decoder accepts the output of the encoder and outputs smoothed data, which can reduce the noise in the input data and improve the overall quality of the data to achieve the effect of feature smoothing.

[0034] Temporal Convolutional Network (TCN) is a special convolutional neural network that can process time series data. TCN can process the entire time series in parallel, making it more efficient than sequence models such as recurrent neural networks (RNN). The use of dilated convolutions allows the model to accommodate both deep networks and long-term valid historical information. This is achieved by using an exponentially large receptive field that can capture information over long time steps. For example, for a given photovoltaic power generation parameter of One-dimensional sequence of and filter , the elements in the sequence , dilated convolution operation The definition is as follows: ,in is the expansion factor, is the filter size, Represents the direction of the past. TCN is constructed using stacked units of one-dimensional convolution with activation functions. In addition, causal convolution is used to ensure that the model does not get future data when predicting future values, maintaining the consistency of time order.

[0035] Figure 3The structure diagram of the temporal convolutional network proposed in the embodiment of the present application is shown, which is divided into an input layer (Input), multiple hidden layers (Hidden Layer) and an output layer (Output), and the expansion factor of the input layer is set ( ), three hidden layers ( ), a fixed step size is introduced between each adjacent filter to expand the coverage of the receptive field. A larger expansion range and filter size k can effectively expand the receptive field. In the convolution operation, the increase of the expansion factor d will exponentially enhance the depth of the network. This ensures that in the effective history, at least one filter can capture the information of each input.

[0036] The first encoder consists of four residual layers, each of which consists of two 1-D CNN layers; The first decoder consists of two dense layers using the LeakyReLU activation function to introduce a negative slope.

[0037] The first encoder includes four residual layers, each of which consists of two one-dimensional convolutional layers (1-D CNN). The residual layer captures the long-term dependencies in the time series data through convolution operations and compresses the high-dimensional input data into low-dimensional latent representations. The Flatten layer flattens the multi-dimensional feature map output by the encoder into a one-dimensional vector so that the decoder can process it. The first decoder includes two dense layers (Dense), using the LeakyReLU activation function to introduce a negative slope to improve the flexibility of the algorithm, restore the low-dimensional latent representation to high-dimensional output data, and generate smoothed data.

[0038] The input data passes through the encoder module, and feature extraction and compression are performed through four residual layers. The output of the encoder is flattened into a one-dimensional vector through the flatten layer, and the one-dimensional vector is input into the decoder module. The data is reconstructed through two dense layers and the LeakyReLU activation function to generate smoothed output data.

[0039] Figure 4 The structure diagram of the temporal convolutional autoencoder proposed in the embodiment of the present application is shown. The input layer is the starting point of the entire model. The encoder consists of four residual blocks, and the expansion rates of each residual block are 1, 2, 4 and 8 respectively. The last layer is a flattening layer, which flattens the output of the encoder into a one-dimensional vector. The decoder consists of three dense layers and a LeakyReLU activation function. The last layer is the output layer, which represents the final output of the model. The parameters of the temporal convolutional autoencoder are shown in Table 1: Table 1

[0040] In data smoothing processing, including: Select the window size and slide the window point by point along the time axis. A new window is generated every time it moves one time unit. Take the average of the initial w data points, repeat w times, and generate supplementary data at the starting position; Average the last w-1 data points and repeat w-1 times to create supplementary data for the end positions.

[0041] The data smoothing process also includes sliding window data construction and mean filling. Through Pearson correlation analysis, six features that are highly correlated with the target variable are screened out. Then, the sliding window is used to split the time series data into subsequences of fixed length. Figure 5 The schematic diagram of the structure of the sliding window proposed in the embodiment of the present application is shown. Assuming that RDT is a one-dimensional time series data, when smoothing the w-th data point of the one-dimensional feature data RDT, arrive The 24 sample data point sequence is used as the input of the encoder layer, and then the sliding window Move forward one step to rebuild . In mean filling, for one-dimensional features, the mean filling method is used to supplement the starting and ending positions of the data. The initial w data points are averaged and repeated w times to generate supplementary data for the starting position. Similarly, the last w-1 data points are averaged and repeated w-1 times to create supplementary data for the end position. This method ensures that the feature samples are aligned with their respective labels. The mean filling method solves the lag problem of reconstructed data. In addition, the smoothing effect of the most recent samples is enhanced. Therefore, the reconstructed feature samples are more robust and can extract key feature information to a greater extent.

[0042] Figure 6 The TCN-AE data smoothing effect diagram proposed in the embodiment of the present application is shown. The six characteristic data selected by the Pearson correlation analysis in the photovoltaic data set are sequentially input into the TCN-AE for data smoothing according to the above method. The whole processing process takes 307s. The original data of three characteristic data of temperature (Wavelet Transform Coefficient, WTC), relative humidity (Derivative of Humidity Ratio, DHR) and global horizontal radiation (Global Horizontal Irradiance, GHR) are listed in the figure and compared with the data after smoothing by TCN-AE. It can be seen that the data smoothing effectively eliminates the noise part of the original data while retaining the potential trend information. After the data is smoothed, the decoded data is respectively input into the encoder and decoder components of the information model to predict the photovoltaic power generation.

[0043] Step S200: Use the LSTM-based operation cost analysis model to perform photovoltaic data dimension analysis on the processed photovoltaic data set to obtain the photovoltaic operation and maintenance cost.

[0044] After obtaining the processed PV data set, select features that are closely related to the operation and maintenance costs or create new features, build an LSTM model and train the LSTM model using the preprocessed and feature-engineered data. Use the reserved test set to evaluate the performance of the model, focusing on the prediction accuracy and stability of the model. If the model performs poorly, it may be necessary to return to the data preparation or feature engineering stage, adjust the data processing method or add / delete certain features, and then retrain the model. Once the model training is completed and passes the evaluation, it can be applied to actual scenarios to estimate the costs of different operation and maintenance solutions based on current and predicted operation and maintenance data.

[0045] The LSTM-based activity-based cost analysis model includes memory units and gate structures. The memory units are used to transmit and store information throughout the sequence and maintain long-term memory. The gate structure includes input gate, forget gate, and output gate; Input gate, used to add information to the memory cell, controls the inflow of information through the sigmoid activation function; The forget gate is used to delete information from the memory unit and use the sigmoid activation function to control the discard of information; The output gate is used to determine the information output in the memory unit and uses the sigmoid activation function to control the outflow of information.

[0046] The memory cell is the core of LSTM, which can transfer and store information throughout the sequence and maintain long-term memory. It stores information through a continuous state, which can be transferred between time steps and is not affected by gradient vanishing or gradient exploding. The input gate adds new information to the memory cell and uses a sigmoid activation function to determine which information needs to be added to the memory cell. The output range of the sigmoid function is between 0 and 1, which controls the inflow of information. The forget gate deletes old information from the memory cell and also uses a sigmoid activation function to determine which information needs to be deleted from the memory cell. The output of the sigmoid function determines the degree of information discarded. The output gate determines how the information in the memory cell is output, and a sigmoid activation function is used to control how the information in the memory cell is output to the next time step. In addition, a tanh activation function is used to adjust the output range so that it is between -1 and 1.

[0047] Step S300: Use the photovoltaic power generation prediction model to predict the power generation of the processed photovoltaic data set to obtain the power generation of the power station equipment.

[0048] The TCN-AE-Informer model is selected as the photovoltaic power generation prediction model, in which the time series convolutional network is used to capture local dependencies in the time series, the autoencoder is used to learn efficient representation of data, and the Informer is an efficient self-attention mechanism model, which is particularly suitable for processing long sequence data. Combining these three technologies, the photovoltaic power generation prediction model can effectively predict photovoltaic power generation. The TCN-AE-Informer model is trained using the preprocessed photovoltaic dataset. During the training process, the model learns how to predict future power generation from historical power generation data. After training, the model is used to predict power generation on the processed photovoltaic dataset and used as input for the subsequent alarm evaluation model.

[0049] The photovoltaic power generation prediction model includes a second encoder and a second decoder; The second encoder includes a multi-head sparse self-attention mechanism and a distillation algorithm. The calculation formula of the second encoder is: ,in , , is the matrix obtained by performing different linear transformations on the input vector, is obtained by probabilistic sparse method. The resulting matrix, are the dimensions of the query vector and key vector, is the normalized activation function; The second decoder consists of two identical multi-head attention layers. The calculation formula of the second decoder is: ,in Indicates the second decoder An input sequence, Mark is The start marker of a sequence, is a placeholder. It is a splicing operation.

[0050] The Informer model is an advanced deep learning architecture suitable for dealing with long-sequence time series prediction problems. Figure 7The Transformer-based model structure proposed in the embodiment of the present application is shown, including an encoder and a decoder. The encoder includes multiple multi-head attention layers and ProbSparse self-attention layers. The multi-head attention layer processes the input sequence, extracts features and generates feature maps. The output of the encoder is a feature map after multi-layer processing. The decoder also includes a multi-head attention layer and a Masked Multi-head Attention layer. The Masked Multi-head Attention layer is used to process the input sequence of the decoder to ensure that future tokens are not seen when generating the sequence. The output of the decoder is a feature map after multi-layer processing. The output feature maps of the encoder and decoder are connected and sent to the fully connected layer. The fully connected layer converts the feature map into the final output sequence. The final output sequence represents the prediction result of the model.

[0051] The multi-head sparse self-attention mechanism in the second encoder reduces the computational complexity by introducing sparsity while maintaining the effectiveness of the model. The calculation formula of the second encoder is: ,in , , is the matrix obtained by performing different linear transformations on the input vector, is obtained by probabilistic sparse method. The resulting matrix, are the dimensions of the query vector and key vector, is the normalized activation function.

[0052] The distillation algorithm reduces the number of network parameters and highlights important features, making the model more efficient and accurate. The calculation formula of the distillation algorithm is: ,in is the attention module of the multi-head sparse self-attention mechanism, represents the convolution operation on the time series, ELU is the activation function, and the distillation algorithm adds a pooling layer with a step size of 2, which increases the memory utilization to ,in The value of is extremely small.

[0053] The second decoder consists of two identical multi-head attention layers to generate predictions for future time steps. This improves the model’s prediction performance by alleviating the slowdown problem of long-term predictions. The calculation formula for the second decoder is: , Indicates the decoder An input sequence, Mark is The start marker of a sequence, is a placeholder, indicating The target sequence of a sequence is composed of the aforementioned 0 values. The decoder uses the intermediate results generated by the encoder to perform multi-head attention operations, adjusts the dimension of the output data through the fully connected layer, and finally generates the predicted output result.

[0054] Step S400: input the power generation of the power station equipment into the Transformer-based alarm evaluation model for data analysis to obtain the power generation saved per unit time.

[0055] Based on the predicted power generation, the alarm evaluation model can more accurately assess the impact of different alarm events on actual power generation, helping operation and maintenance personnel to take timely measures to reduce losses. Figure 8 The structure diagram of the Transformer model proposed in the embodiment of the present application is shown. The input data is first embedded in the input, the input symbol is converted into a vector representation, the vector is added to the position code, and the position information of each symbol in the input sequence is retained. The encoder consists of multiple identical layers, and the multi-head self-attention captures the dependencies between different positions in the input sequence. The feedforward neural network performs a nonlinear transformation on the output of the attention mechanism. The output is added to the input through a residual connection (Add&Norm) and normalized to maintain information transmission.

[0056] The decoder consists of multiple identical layers, where masked multi-head self-attention captures dependencies between different positions in the output sequence while preventing the model from seeing future outputs, and multi-head attention captures dependencies between the input sequence and the output sequence. The feedforward neural network performs a nonlinear transformation on the output of the attention mechanism. The output of the decoder is output embedded and then passes through a linear layer and a softmax function to generate the final output probability distribution. Position encoding adds position information to each symbol in the input and output sequences so that the model can distinguish symbols at different positions.

[0057] The Transformer-based alarm evaluation model includes a third encoder, a third decoder, and a self-attention mechanism; The third encoder includes multiple identical stacked layers, each of which contains a multi-head self-attention mechanism and a fully connected feedforward network. The outputs of the multi-head self-attention mechanism and the fully connected feedforward network are processed by residual connection and layer normalization. The third decoder also includes multiple identical stacked layers, each of which includes a masked multi-head self-attention mechanism, a multi-head self-attention mechanism, and a fully connected feedforward network; The self-attention mechanism uses an attention function to map query and key-value pairs to an output, which is calculated as a weighted sum of values, with the weights calculated by a compatibility function between the query and the corresponding key.

[0058] The third encoder converts the original input data into a high-dimensional abstract representation that can be understood by the computer. The third encoder consists of multiple (for example, six) identical stacked layers, each of which contains two sublayers, namely the multi-head self-attention mechanism and the fully connected feedforward network. The multi-head self-attention mechanism captures the dependencies between different positions in the input sequence through multiple parallel self-attention heads. The fully connected feedforward network, as a simple linear transformation layer, can enhance the expressiveness of the model. The output of each sublayer will be connected through residual connection (LayerNorm(x+Sublayer(x)), where Sublayer(x) is the function implemented by the sublayer itself) to alleviate the gradient vanishing problem, while maintaining the stability and consistency of the output through layer normalization.

[0059] The third decoder uses the high-dimensional representation information generated by the encoder to generate the target sequence. It consists of multiple (for example, six) identical stacked layers, each of which contains three sublayers, namely, the masked multi-head self-attention mechanism, the multi-head self-attention mechanism, and the fully connected feedforward network. The masked multi-head self-attention mechanism is similar to the multi-head self-attention mechanism in the encoder, but is masked to ensure that when predicting a certain position, it can only rely on the information before that position. The multi-head self-attention mechanism performs multi-head attention on the output of the encoder so that the decoder can make full use of the high-dimensional representation generated by the encoder. The fully connected feedforward network is similar to the fully connected feedforward network in the encoder and is used to enhance the expressiveness of the model. Like the encoder, the output of each sublayer is also processed by residual connection and layer normalization.

[0060] The query and a set of key-value pairs are mapped to the output through the self-attention mechanism, capturing the dependencies between different positions in the sequence. The attention function maps the query, key, and value to the output. The output is the weighted sum of the values, and the weight is calculated by the compatibility function of the query and the corresponding key. Through multiple parallel attention heads, each head independently calculates the attention weight, and then the output of all heads is spliced ​​together and passed through a linear transformation layer to form the final output. In the same sequence, the query, key, and value of each position come from the same sequence, allowing the model to calculate the degree of association between the position and other positions at each position, and weight the information in the combined sequence according to these association degrees.

[0061] Step S500: Calculate the operation and maintenance benefits based on the power generation recovered per unit time and the photovoltaic operation and maintenance costs to evaluate the overall operation benefits.

[0062] Operation and maintenance income = recovery income per unit time * evaluation time - photovoltaic operation and maintenance cost, where the recovery income per unit time is based on the power generation recovered per unit time.

[0063] Fig. 9The flowchart of photovoltaic prediction, alarm assessment and operation cost analysis proposed in the embodiment of the present application is shown. First, the power generation of the photovoltaic power station is predicted, and historical power generation data, weather data, etc. are input and collected, and then the data is cleaned and sorted to ensure the quality and consistency of the data. The pre-processed data is trained using the Informer model, and the trained model is used to predict the power generation in the future based on the input real-time data.

[0064] Then predict the power loss that may be caused by equipment failure, collect equipment operation data, alarm information, etc., clean and organize the data to ensure data quality and consistency, and use the Transformer model to train the pre-processed data. The Transformer model performs well in processing sequence data. Using the trained model, based on the input equipment status and alarm information, predict the power loss that may be caused by equipment failure.

[0065] Operational cost analysis predicts operation and maintenance costs, collects historical operation and maintenance data, equipment maintenance records, etc., cleans and organizes data to ensure data quality and consistency, uses LSTM model to train preprocessed data, and uses the trained model to predict operation and maintenance costs based on input operation and maintenance requirements and equipment status.

[0066] The operation benefit assessment objective comprehensively evaluates the economic feasibility of the operation and maintenance plan, and calculates the total benefit of the operation and maintenance plan by combining the results of photovoltaic forecasting, alarm assessment and operation cost analysis. The specific calculation formula includes: operation and maintenance benefit = lost electricity × unit electricity price - operation and maintenance cost.

[0067] In summary, the operation cost analysis model is used to convert the equipment that needs to be repaired and the alarm information into operation and maintenance costs. The photovoltaic power generation prediction model is used to convert the equipment status, weather conditions, etc. into the predicted equipment power generation. Then, the alarm evaluation model is used to convert the equipment that needs to be repaired, the alarm information, and the predicted power generation into the power generation recovered per unit time. Finally, the overall benefit of the operation is evaluated.

[0068] Compared with the prior art, the embodiments of the present application have the following beneficial effects: First, through cost forecasting, it helps power station managers formulate reasonable budgets and maintenance plans to reduce operation and maintenance costs.

[0069] Second, by predicting the amount of lost electricity, it helps power station managers provide equipment repair priorities and increase operation and maintenance benefits.

[0070] Second, by evaluating the economic feasibility of the overall operation and maintenance plan, we help power plant management formulate the best operation and maintenance plan, change the chaotic and aimless operation and maintenance methods of the power plant, and make the power plant operation and maintenance efficient and simple.

[0071] A possible implementation of an activity economics evaluation system based on big data analysis is given below, which is used to execute the various execution steps and corresponding technical effects of the activity economics evaluation method shown in the above embodiments and possible implementations. The system includes: A data smoothing module is used to perform data smoothing on the photovoltaic data set using a time series convolutional autoencoder to obtain a processed photovoltaic data set; The first analysis module is used to use the LSTM-based operation cost analysis model to perform photovoltaic data dimension analysis on the processed photovoltaic data set to obtain the photovoltaic operation and maintenance cost; A power generation prediction module is used to use a photovoltaic power generation prediction model to predict the power generation of the processed photovoltaic data set to obtain the power generation of the power station equipment; The second module is used to input the power generation of power station equipment into the Transformer-based alarm evaluation model for data analysis to obtain the power generation saved per unit time; The calculation module is used to calculate the operation and maintenance benefits based on the power generation recovered per unit time and the photovoltaic operation and maintenance costs to evaluate the overall operation benefits.

[0072] The above description is only a preferred embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present application should be included in the protection scope of the present application.

Claims

1. A method for evaluating the economic performance of an operation based on big data analysis, characterized in that: The method comprises: The photovoltaic data set is smoothed using a time series convolutional autoencoder to obtain a processed photovoltaic data set, which includes historical operation and maintenance records, alarm information, equipment status, environmental factors, labor costs, transportation costs, spare parts, and tool costs. The LSTM-based operation cost analysis model is used to analyze the photovoltaic data dimension of the processed photovoltaic data set to obtain the photovoltaic operation and maintenance cost; Use the photovoltaic power generation prediction model to predict the power generation of the processed photovoltaic data set to obtain the power generation of the power station equipment; Input the power generation of power station equipment into the Transformer-based alarm evaluation model for data analysis to obtain the power generation saved per unit time; The operation and maintenance benefits are calculated based on the power generation recovered per unit time and the photovoltaic operation and maintenance costs to evaluate the overall benefits of the operation.

2. The method for evaluating the economic efficiency of an operation according to claim 1, wherein: The temporal convolutional autoencoder includes an autoencoder and a temporal convolutional network. The autoencoder learns the optimal representation of the input features of the photovoltaic dataset by minimizing the reconstruction error through self-supervised learning. The temporal convolutional network is used to utilize the parallel processing capabilities of convolutional neural networks for data processing, uses dilated convolutions to capture long-term dependencies, and learns temporal features by performing convolution operations on a one-dimensional time axis.

3. The method for evaluating the economic efficiency of an operation according to claim 2, wherein: The autoencoder includes a first encoder, a first decoder and a flatten layer. The first encoder is used to compress high-dimensional input data into a low-dimensional latent representation, the first decoder is used to restore the low-dimensional latent representation to high-dimensional output data, and the flatten layer is used to flatten the output of the first encoder and then input it into the first decoder.

4. The method for evaluating the economic efficiency of an operation according to claim 3, wherein: The first encoder consists of four residual layers, each of which consists of two 1-D CNN layers; The first decoder consists of two dense layers using the LeakyReLU activation function to introduce a negative slope.

5. The method for evaluating the economic efficiency of an operation according to claim 1, wherein: In data smoothing processing, including: Select the window size and slide the window point by point along the time axis. A new window is generated every time it moves one time unit. Take the average of the initial w data points, repeat w times, and generate supplementary data at the starting position; Average the last w-1 data points and repeat w-1 times to create supplementary data for the end positions.

6. The method for evaluating the economic efficiency of an operation according to claim 1, wherein: The LSTM-based activity-based cost analysis model includes memory units and gate structures. The memory units are used to transmit and store information throughout the sequence and maintain long-term memory. The gate structure includes input gate, forget gate, and output gate; Input gate, used to add information to the memory cell, controls the inflow of information through the sigmoid activation function; The forget gate is used to delete information from the memory unit and use the sigmoid activation function to control the discard of information; The output gate is used to determine the information output in the memory unit and uses the sigmoid activation function to control the outflow of information.

7. The method for evaluating activity economics according to claim 1, wherein: The photovoltaic power generation prediction model includes a second encoder and a second decoder; The second encoder includes a multi-head sparse self-attention mechanism and a distillation algorithm. The calculation formula of the second encoder is: ,in , , is the matrix obtained by performing different linear transformations on the input vector, is obtained by probabilistic sparse method. The resulting matrix, are the dimensions of the query vector and key vector, is the normalized activation function; The second decoder consists of two identical multi-head attention layers. The calculation formula of the second decoder is: ,in Indicates the second decoder An input sequence, Mark is The start marker of a sequence, is a placeholder. It is a splicing operation.

8. The method for evaluating activity economics according to claim 1, wherein: The Transformer-based alarm evaluation model includes a third encoder, a third decoder, and a self-attention mechanism; The third encoder includes multiple identical stacked layers, each of which contains a multi-head self-attention mechanism and a fully connected feedforward network. The outputs of the multi-head self-attention mechanism and the fully connected feedforward network are processed by residual connection and layer normalization. The third decoder also includes multiple identical stacked layers, each of which includes a masked multi-head self-attention mechanism, a multi-head self-attention mechanism, and a fully connected feedforward network; The self-attention mechanism uses an attention function to map query and key-value pairs to an output, which is calculated as a weighted sum of values, with the weights calculated by a compatibility function between the query and the corresponding key.

9. The method for evaluating activity economics according to claim 1, wherein: Operation and maintenance income = unit time recovery income * evaluation time - photovoltaic operation and maintenance cost, where the unit time recovery income is based on the power generation recovered per unit time.

10. An operation economic evaluation system based on big data analysis, characterized in that: The system comprises: A data smoothing module is used to perform data smoothing on the photovoltaic data set using a time series convolutional autoencoder to obtain a processed photovoltaic data set; The first analysis module is used to use the LSTM-based operation cost analysis model to perform photovoltaic data dimension analysis on the processed photovoltaic data set to obtain the photovoltaic operation and maintenance cost; A power generation prediction module is used to use a photovoltaic power generation prediction model to predict the power generation of the processed photovoltaic data set to obtain the power generation of the power station equipment; The second module is used to input the power generation of power station equipment into the Transformer-based alarm evaluation model for data analysis to obtain the power generation saved per unit time; The calculation module is used to calculate the operation and maintenance benefits based on the power generation recovered per unit time and the photovoltaic operation and maintenance costs to evaluate the overall operation benefits.

Images