Method for predicting concentration of dissolved gas in transformer oil

Abstract

The invention belongs to the field of power equipment monitoring, and relates to a method for predicting the concentration of dissolved gas in transformer oil, which comprises the following steps of: processing a dissolved gas concentration data set by applying adaptive noise complete set empirical mode decomposition to obtain a plurality of intrinsic mode functions; a hybrid deep learning model based on a time convolutional network, a bidirectional long-short-term memory network and a converter architecture is constructed, the time convolutional network is used for capturing local features, the bidirectional long-short-term memory network is responsible for processing a long-term dependency relationship, and a converter realizes information interaction by using a self-attention mechanism; inputting the training set into a mixed deep learning model for training to obtain a prediction model, and inputting the test set into the prediction model for testing, verification and optimization to obtain a prediction optimization model; and giving a final dissolved gas concentration prediction value by utilizing the prediction optimization model. The advanced signal processing technology and the deep learning algorithm are applied to power equipment maintenance, and technical support can be provided for intelligent power grid construction.

Description

Technical Field

[0001] The invention belongs to the technical field of power equipment status monitoring, and in particular relates to a method for predicting dissolved gas concentration in transformer oil. Background Art

[0002] In modern power systems, transformers are key components connecting power grids of varying voltage levels, and their stability and reliability are crucial. Dissolved gas analysis (DGA) in transformer oil is an important tool for assessing transformer health, as changes in dissolved gas concentration can reveal potential faults within the transformer, such as overheating and discharge.

[0003] Traditionally, DGA relies on periodic sampling of transformer oil and sending it to a laboratory for chromatographic analysis. This method suffers from slow response, high cost, and the inability to monitor in real time. With the advancement of information technology, machine learning and deep learning prediction methods based on historical data have gradually become a research hotspot. However, time series data on dissolved gas concentrations in transformer oil are often nonlinear and non-stationary, posing a challenge to existing prediction methods. Existing prediction models often struggle to accurately capture these complex characteristics, resulting in low prediction accuracy and limiting their effectiveness in practical applications. Therefore, it is particularly necessary to develop a new method that can effectively address these technical issues. Summary of the Invention

[0004] To address the above-mentioned technical problems, the present invention proposes an innovative method for predicting dissolved gas concentration in transformer oil, overcoming the shortcomings of the existing technology. First, the present invention introduces the Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) technique, an improved empirical mode decomposition technique that enhances the stability of the original signal by adding white noise and more effectively separates the different frequency components within the signal. Next, for each intrinsic mode function (IMF) decomposed by CEEMDAN, the present invention designs a deep learning framework that integrates a temporal convolutional network (TCN), a bidirectional long short-term memory network (BiLSTM), and a transformer. TCN excels at capturing local features, BiLSTM can handle long-term dependencies, and the transformer utilizes a self-attention mechanism to achieve efficient information exchange. The combination of these three forms a powerful and flexible prediction model that can extract features from signals at different scales, significantly improving prediction accuracy and robustness. Furthermore, the present invention takes into account the generalization capability and computational efficiency of the prediction model, minimizing resource consumption while ensuring performance. The technical solutions adopted by the present invention are as follows: A method for predicting dissolved gas concentration in transformer oil collects dissolved gas concentration data in transformer oil and generates a time series data set of dissolved gas concentration in transformer oil, comprising the following steps: Applying adaptive noise complete ensemble empirical mode decomposition to process the time series data set of dissolved gas concentration in the transformer oil, a number of intrinsic mode functions with different frequency characteristics are obtained; Build a hybrid deep learning model based on a temporal convolutional network, a bidirectional long short-term memory network, and a transformer architecture. The temporal convolutional network is used to capture local features, the bidirectional long short-term memory network is responsible for processing long-term dependencies, and the transformer uses a self-attention mechanism to achieve information exchange. All intrinsic mode functions are divided into a training set and a test set, the training set is input into a hybrid deep learning model for training to obtain a prediction model for dissolved gas concentration in transformer oil, and the test set is input into the prediction model for dissolved gas concentration in transformer oil for testing, verification and optimization to obtain an optimized prediction model for dissolved gas concentration in transformer oil; Finally, the optimization model for predicting dissolved gas concentration in transformer oil is used to give the final predicted value of dissolved gas concentration in transformer oil.

[0005] Preferably, the adaptive noise complete set empirical mode decomposition is applied to process the time series data set of the dissolved gas concentration in the transformer oil. The specific method is: the time series data set is denoised and smoothed, and the data is decomposed into a number of intrinsic mode functions with physical significance by adding white noise of different phases multiple times and then averaging. Each intrinsic mode function represents the fluctuation characteristics of the original data within a specific frequency range.

[0006] Preferably, each intrinsic mode function is adjusted by an interpolation method so that their time points correspond one to one; the prediction sequence of each intrinsic mode function is padded or trimmed. When the prediction sequence of a certain intrinsic mode function is longer than other sequences, the redundant data points are trimmed from both ends. If the prediction sequence of the intrinsic mode function is short, the last valid value is added to the end of the prediction sequence or linear interpolation is used to supplement the missing data points. After data padding or trimming, the prediction results of each intrinsic mode function have equal length within the same time range.

[0007] Preferably, a K-fold cross-validation method is used to divide the training data into K non-overlapping subsets, which are used in turn as test sets to evaluate the prediction model.

[0008] Preferably, an optimizer and a loss function are configured for the hybrid deep learning model of each intrinsic mode function, and the hybrid deep learning model is compiled; the optimizer uses the Adam optimizer, and the loss function uses the mean square error.

[0009] Preferably, Python language is used for programming, combined with the TensorFlow deep learning framework; Docker container technology is used to encapsulate the learning model; the hybrid deep learning model is deployed to a cloud server, or the hybrid deep learning model is deployed to an edge device.

[0010] Preferably, the method for collecting data on dissolved gas concentration in transformer oil is to regularly extract oil samples from a running transformer and measure the dissolved gas concentration in the transformer oil using a portable or online monitoring device.

[0011] Beneficial effects of the present invention: This method combines adaptive noise complete ensemble empirical mode decomposition (EMD) technology with a temporal convolutional network-bidirectional long short-term memory network-transformer (TCN-BiLSTM-Transformer) hybrid deep learning model. It aims to improve early detection and early warning capabilities for potential transformer internal faults, ensuring the safe and reliable operation of power systems. By applying advanced signal processing techniques and deep learning algorithms to power equipment maintenance, this method not only helps optimize substation operation and maintenance strategies but also provides technical support for the development of smart grids. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings: Figure 1 The figure is a flowchart of the steps of a method for predicting the concentration of dissolved gas in transformer oil according to an embodiment of the present invention. DETAILED DESCRIPTION

[0013] The technical solution of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, but not all of the embodiments.

[0014] like Figure 1 To achieve the above-mentioned purpose, an embodiment of the present invention provides a method for predicting the concentration of dissolved gas in transformer oil, comprising the following steps: 1. Data collection and preparation (transformer DGA raw data).

[0015] Data collection: Transformer oil samples are regularly taken from operating transformers, and the concentrations of dissolved gases (such as carbon monoxide, hydrogen, methane, acetylene, etc.) in the transformer oil are measured using portable or online monitoring devices to form a time series dataset of dissolved gas concentrations in the transformer oil.

[0016] Load Data: Loads a time series dataset containing dissolved gas concentrations in transformer oil from a specified path. In this example, the present invention focuses on carbon monoxide (CO) data.

[0017] Select target variable: Extract carbon monoxide concentration as the target variable for subsequent decomposition and prediction.

[0018] Ensuring the completeness and accuracy of data is crucial for subsequent analysis.

[0019] 2. Time series dataset preprocessing.

[0020] The complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) is used to denoise and smooth the time series data set. By adding white noise of different phases multiple times and then averaging the data, the data is decomposed into several physically meaningful intrinsic mode function (IMF) components. Each IMF component represents the fluctuation characteristics of the original data within a specific frequency range.

[0021] A CEEMDAN function is constructed to decompose the time series dataset and generate multiple intrinsic mode functions. CEEMDAN (Complete Ensemble Empirical Mode Decomposition with Adaptive Noise) is an advanced signal processing technique particularly suitable for nonlinear and non-stationary time series data. The CEEMDAN function enhances the stability and accuracy of data decomposition by introducing adaptive white noise. Specifically, the CEEMDAN function in this embodiment of the present invention performs the following operations: ① Initialization: First, add a certain amount of independent and identically distributed Gaussian white noise to the original signal x(t) to form a new set of noisy signals.

[0022] ②EMD decomposition: For each noisy signal, use the EMD algorithm to decompose it to obtain a series of IMF components and the residual term r(t).

[0023] ③Averaging: Calculate the mean of the IMF components of the corresponding order of all noisy signals as one of the IMF components of the final output.

[0024] ④ Iterative update: Repeat the above steps with the remaining part as the new input signal until the preset conditions are met, such as reaching the maximum number of iterations MaxIter or no valid IMF components can be extracted.

[0025] Intrinsic Mode Function (IMF) refers to a specific type of signal component obtained after a screening process. Each IMF component is an oscillation mode extracted from the original signal, and each represents a different frequency component or time scale characteristic of the original signal.

[0026] Calculate the maximum mutual information (MIC) coefficient between each IMF component and the original power data, and retain the top 7 items in descending order according to the size of the MIC coefficient.

[0027] Normalization: For each IMF component, apply the Min-Max Normalization method to scale its value to the interval [0, 1] to facilitate model training, and record the minimum and maximum values of each IMF component for subsequent denormalization operations.

[0028] 3. Use training set data to build a hybrid deep learning model.

[0029] A hybrid deep learning model based on the TCN-BiLSTM-Transformer architecture is defined. This not only captures the dynamic characteristics at different frequencies in a more detailed manner, but also effectively reduces the complexity and computational cost of the learning model.

[0030] This embodiment of the present invention utilizes a hybrid deep learning architecture that combines a temporal convolutional network, a bidirectional long short-term memory network, and a transformer. This architecture is designed to process nonlinear and nonstationary time series data, aiming to capture local features, long-term dependencies, and efficient information exchange. To better understand the connections between the modules, this embodiment of the present invention presents the entire process in block diagram form, with detailed explanations of the functions of each component and its input and output parameters.

[0031] Functions and parameter configurations of each module of the hybrid deep learning model: 3.1. Input layer

[0032] Input layer input: accepts data of shape (seq_length, 1), where seq_length is the length of the time series.

[0033] Input layer output: data with the original shape unchanged is passed directly to the next layer.

[0034] 3.2. Custom TCN layer

[0035] Implementation: TCN layers are constructed by stacking multiple one-dimensional convolutional layers, each using dilated convolution. Dilated convolutions allow for a wider receptive field without adding additional parameters, thus covering information over a longer timeframe. Furthermore, TCN layers utilize residual connections, or skip connections, which append the input of one layer directly to the output of the subsequent layer. This helps alleviate the vanishing gradient problem and accelerates training.

[0036] Parameter configuration: Number of filters: determines the number of channels or feature dimensions after each convolution layer.

[0037] KernelSize: Specifies the width of the convolution kernel. Generally, a smaller value such as 3 is selected.

[0038] Dilation Rate: controls the spacing between elements in the convolution operation, increasing layer by layer to expand the receptive field.

[0039] Padding: Use "causal" padding to ensure causality, that is, future moments will not affect the prediction of the current moment.

[0040] Activation Function: ReLU activation function is usually used to introduce nonlinearity.

[0041] Batch Normalization: used to stabilize and accelerate the training process.

[0042] TCN layer input: receives data from the input layer, with a shape of (batch_size, seq_length, 1).

[0043] TCN layer output: Generates a new feature representation, and the shape becomes (batch_size, seq_length, filters), where filters represents the feature dimension obtained after convolution.

[0044] 3.3. Custom BiLSTM layer

[0045] A BiLSTM layer is added after the TCN layer to capture long-term dependencies in the sequence. The BiLSTM layer consists of two LSTMs running in opposite directions: one processes the forward sequence and the other processes the reverse sequence. The results of the two are then merged, allowing for the simultaneous consideration of past and future contextual information.

[0046] BiLSTM layer input: Feature representation obtained from the TCN layer, with a shape of (batch_size, seq_length, filters).

[0047] The BiLSTM layer outputs a hidden state sequence containing bidirectional information, also of shape (batch_size, seq_length, hidden_units), where hidden_units is the state dimension inside the BiLSTM layer.

[0048] 3.4. Custom position encoding layer.

[0049] A positional encoding layer based on sine and cosine functions is introduced to add position information to each element in the sequence. This is crucial for the Transformer architecture, which lacks the ability to understand order. The positional encoding layer helps the model distinguish between elements with the same content that appear in different positions.

[0050] Positional encoding layer input: the output of the BiLSTM layer, with a shape of (batch_size, seq_length, hidden_units).

[0051] Position encoding layer output: feature representation with position information, while the shape remains unchanged.

[0052] 3.5. Custom Transformer encoder block.

[0053] It includes a multi-head self-attention mechanism (MultiHeadAttention), a feedforward neural network (FFN), residual connections, and layer normalization. These components work together to achieve efficient information exchange and can handle long dependencies.

[0054] Transformer encoder block input: feature representation after the positional encoding layer, with shape (batch_size, seq_length, hidden_units).

[0055] Transformer encoder block output: enhanced feature representation, shape is still (batch_size, seq_length, hidden_units).

[0056] 3.6. Output through the fully connected layer.

[0057] Finally, the prediction result is output through the fully connected layer (DenseLayer), taking only the result of the last time step as the final output. The fully connected layer is responsible for mapping high-level abstract features back to the space of the target variable.

[0058] Fully connected layer input: the output of the Transformer encoder block, with shape (batch_size, seq_length, hidden_units).

[0059] Fully connected layer output: predicted dissolved gas concentration value, the shape is simplified to (batch_size, 1) or (batch_size, prediction_steps) for multi-step prediction.

[0060] The modules in this part are connected in series in sequence, that is, several normalized IMF components will enter the input layer, TCN layer, BiLSTM layer, position encoding layer, and Transformer encoder block in sequence until they reach the fully connected layer to output the predicted value of the dissolved gas concentration in the transformer oil in the future time period.

[0061] 4. Data partitioning: Divide the data set including all IMF components into training set and test set. Usually 80% of the data is used as the training set for training, and the remaining 20% of the data is used as the test set for testing.

[0062] 5. Use the test set data to train the hybrid deep learning model to generate a prediction model for the dissolved gas concentration in transformer oil; use the test set to verify and optimize the prediction model for the dissolved gas concentration in transformer oil to obtain an optimized prediction model for the dissolved gas concentration in transformer oil.

[0063] Learning model compilation: Configure the optimizer and loss function for the learning model of each IMF component and compile the learning model.

[0064] Optimizer: It is recommended to use the Adam optimizer, which combines the advantages of momentum gradient descent and RMSProp and has better convergence speed and stability.

[0065] Loss function: Considering that the goal of the prediction task is a regression problem, Mean Squared Error (MSE) is a suitable choice because it can effectively measure the difference between the predicted value and the true value.

[0066] Learning and training: Train the corresponding learning model for each IMF component separately, iterate a certain number of times (for example, 500 epochs), and monitor the performance changes during the training process.

[0067] Prediction and denormalization: The trained learning model is used to generate a prediction model to predict the dissolved gas concentration in transformer oil in the future time period, and then the prediction results are converted from the normalized range back to the original scale.

[0068] Training Process: Leverage a large, annotated historical dataset to train the learning model for each IMF component. During training, monitor performance metrics (such as MSE, RMSE, and MAE) on the test set to prevent overfitting and adjust hyperparameters promptly. Early stopping can also be used to automatically terminate training rounds that no longer improve, saving computing resources.

[0069] Once the learning model is trained, it can be used to predict the dissolved gas concentration in transformer oil for future time periods. Simply input the latest IMF component, and the prediction model will calculate the corresponding prediction results. It is important to note that because different IMF components correspond to different time scales, the final prediction result should be a comprehensive consideration of all predicted values.

[0070] 6. Reconstruction of prediction results and verification of prediction model accuracy.

[0071] 6.1. Reconstruct prediction results.

[0072] After completing the prediction for each IMF component, these predictions need to be integrated to obtain the final prediction result of the dissolved gas concentration in the transformer oil. Because different IMF components represent different frequency characteristics of the original signal, their time scales may vary. To ensure that the prediction results of each IMF component are consistent in length, the following measures can be taken: Time alignment: For all IMF component forecasts, ensure that their corresponding timestamps match exactly. If there are time deviations between the forecast results of some IMF components, they can be adjusted through interpolation and other methods to make their time points correspond one-to-one.

[0073] Data padding or trimming: When the forecast sequence of a certain IMF component is longer than other sequences, you can consider trimming the excess data points from both ends. Conversely, if the forecast sequence is shorter, you can add the last valid value at the end of the sequence (i.e., forward padding) or use linear interpolation to supplement the missing data points.

[0074] After the above processing, the IMF forecast results will have equal length within the same time frame, and they can now be directly added together to obtain the final forecast value. It is worth noting that in actual operation, the most appropriate alignment and synchronization strategy should be selected according to the specific situation to ensure the accuracy and reliability of the forecast results.

[0075] 6.2. Calculate evaluation indicators.

[0076] In order to evaluate the performance of the prediction model, it is necessary to calculate the error between the predicted value and the true value. The evaluation indicators used here include root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE). Specifically: Benchmark: The true values used to calculate evaluation metrics should come from the test set. This means that a portion of the data held back during training does not participate in the learning process of the model and is used only to test the generalization ability of the learned model. This allows for a more objective assessment of the model's prediction performance on future data.

[0077] Denormalization and Overlay: Because each IMF component is independently normalized during preprocessing, the predicted values of each IMF component in the test set must be denormalized to their original scale before calculating the evaluation index. These values must then be overlaid in the original decomposition order to restore the complete prediction sequence. This ensures that the predicted values are of the same order of magnitude as the true values, allowing for effective comparison.

[0078] In summary, the correct approach is to use the IMF components in the test set for prediction, then denormalize and overlay these predictions, ultimately comparing them with the actual measurements in the test set as the basis for calculating various evaluation metrics. This not only adheres to the principles of scientific verification but also provides more accurate feedback on the performance of the learning model.

[0079] Visualization: Plot a comparison chart of actual values and predicted values to intuitively show the performance of the model and help understand the accuracy of the prediction.

[0080] Compare the predicted results with the actual measured values to evaluate the performance of the prediction model. If significant deviations are found, retrain the learning model or adjust the preprocessing process. At the same time, establish a long-term data tracking mechanism to continuously accumulate new samples for learning model training to ensure that it is always in optimal working condition.

[0081] 7. Performance optimization and improvement (hyperparameter adjustment and cross-validation).

[0082] To ensure the effectiveness and efficiency of hybrid deep learning models, ensuring optimal predictive performance, good generalization, and stability, and further improving the generalization and robustness of the predictive model, K-fold cross-validation can be employed. This involves dividing the training data into K non-overlapping subsets, which are rotated and evaluated as test sets. This not only yields more reliable performance estimates but also helps identify potential issues, such as whether the predictive model is overly sensitive to a particular data distribution. Based on the cross-validation results, various hyperparameters can be fine-tuned until the optimal configuration is found. The following details the specific implementation methods for hyperparameter adjustment and cross-validation.

[0083] 7.1. Hyperparameter Tuning

[0084] The choice of hyperparameters has a crucial impact on the performance of the prediction model. When building and training the TCN-BiLSTM-Transformer hybrid deep learning model, the hyperparameters in the following key steps require special attention: 7.1.1, TCN layer

[0085] The number of layers is determined by the complexity of the IMF components. For high-frequency IMF components, fewer layers (e.g., 2-3 layers) can be used, while for low-frequency components, more layers (e.g., 4-6 layers) may be required to fully extract long-term dependencies.

[0086] Number of filters: Optimal values can be explored within a certain range using grid search or random search. The initial value can be set to 64 and gradually doubled as the number of layers increases (e.g., 64 filters in the first layer, 128 in the second layer, etc.) to enhance feature representation capabilities.

[0087] Kernel size: You can try different sizes to see how they affect model performance. It’s usually set to 3, which is a compromise between being too wide to cause information loss and too narrow to affect the receptive field.

[0088] Dilation rate: The dilation rate growth pattern can be adjusted based on experimental results. Start from 1 and increase layer by layer (e.g. 1, 2, 4, ...) to expand the receptive field without increasing the number of parameters.

[0089] 7.1.2. BiLSTM layer

[0090] Hidden Units: This represents the dimension of the LSTM's internal state. A higher number of hidden units helps capture more complex sequential patterns and dependencies, but it can also lead to overfitting. Therefore, the optimal number of hidden units should be determined based on test set performance. It's recommended to start with 128 or 256, depending on available computing resources and experimental results.

[0091] Layers: Increasing the number of BiLSTM layers may improve the learning model's ability to capture long-term dependencies, but it also increases computational complexity. It's recommended to start with a single layer and gradually increase the number of layers until performance no longer significantly improves. Generally, a single BiLSTM layer is sufficient; however, for particularly complex IMF components, consider increasing the number to two layers.

[0092] 7.1.3. Transformer encoder block.

[0093] The number of heads in the MultiHeadAttention mechanism (MultiHeadAttention) determines the number of different subspaces processed in parallel. A larger number of heads can enhance the information exchange capabilities of the learning model, but it is also important to consider the consumption of computing resources. It is recommended to start with 8 or 16 heads, which can provide sufficient parallel processing capacity while maintaining computational efficiency.

[0094] Intermediate Dimension of Feedforward Neural Networks (FFNs): Larger intermediate layer dimensions provide greater expressiveness but also carry the risk of overfitting. This should be appropriately set based on specific task requirements. A common practice is to set it to four times the hidden state dimension, which improves expressiveness without excessively increasing the number of parameters.

[0095] 7.1.4. Batch size (BatchSize).

[0096] Batch size directly affects the quality of gradient estimation and memory usage. Smaller batch sizes can introduce more noise and help escape local minima, while larger batch sizes tend to provide more stable gradient updates. Choose an appropriate batch size based on hardware conditions and experimental results, typically between 32 and 128.

[0097] 7.1.5. Learning Rate

[0098] The learning rate is one of the most sensitive hyperparameters in the optimizer. A high value can lead to unstable training, while a low value can slow convergence. You can use a learning rate decay strategy or a dynamic adjustment method such as the Adam optimizer's built-in learning rate scheduling mechanism. An initial value of 0.001 is a reasonable choice, and you can then adjust the learning rate using a decay strategy based on test set performance.

[0099] 7.2. Cross-validation.

[0100] To ensure the objectivity and reliability of the prediction model evaluation, the present invention adopts a K-fold cross validation strategy. The specific steps are as follows: 7.2.1. Data partitioning: Randomly divide the entire dataset into K mutually disjoint subsets (folds), each containing approximately the same number of data points.

[0101] 7.2.2. Training and validation loop: For the kth iteration (k=1,2,...,K), the kth subset is selected as the test set, and the remaining K-1 subsets are combined into the training set.

[0102] Train the learning model on the training set and evaluate the prediction model performance using the test set. Record the evaluation metrics (such as RMSE, MAE, and MAPE) for this iteration.

[0103] 7.2.3 Average Performance Evaluation: After completing all K iterations, calculate the average value and standard deviation of each evaluation metric. This provides a stable estimate of the generalization ability of the prediction model.

[0104] 7.2.4. Final prediction model training: Once the optimal hyperparameter combination is determined, the prediction model is retrained using the complete training data to obtain the final prediction model version for practical application.

[0105] Input parameters: For cross-validation, the main input parameters include the original dataset, K value (usually 5 or 10), prediction model architecture definition, and hyperparameter configuration.

[0106] Output: The main output of cross-validation is the evaluation metric score for each fold, along with statistical summaries of these scores (such as mean and standard deviation). In addition, a set of validated optimal hyperparameter settings is generated for subsequent full-data training.

[0107] By systematically performing hyperparameter adjustment and cross-validation, not only can the prediction accuracy of the prediction model be improved, but also its stable performance under different data distributions can be ensured, thus providing a solid guarantee for the accurate prediction of dissolved gas concentration in transformer oil.

[0108] Continuous Updates: As new data accumulates, the forecast model is regularly retrained to ensure it is always in top condition and accurately reflects the latest trends and developments.

[0109] Through the above steps, the present invention constructs an efficient and accurate hybrid deep learning model and optimized prediction model for predicting dissolved gas concentration in transformer oil. This model not only captures the complex characteristics of time series data at different frequency scales but also exhibits good generalization and stability, providing a solid foundation for practical applications. The model not only achieves excellent prediction accuracy but also maintains efficient computational performance and good generalization.

[0110] The prediction model supports an end-to-end learning model, requiring no additional manual feature engineering or preprocessing steps, and directly learns useful feature representations from the original IMFs. The prediction model has good scalability and portability. To ensure the scalability and portability of the model, the embodiments of the present invention take the following measures: Programming language and framework: Python language is used for programming, combined with TensorFlow deep learning framework.

[0111] Containerized deployment: Use container technologies like Docker to encapsulate models and ensure they can run seamlessly on different operating systems and hardware platforms. Containerization can isolate environment variables, avoid dependency conflicts, and facilitate deployment and management.

[0112] Cloud service integration: Deploy the model to cloud servers such as Alibaba Cloud, AWS, or Google Cloud Platform, leveraging the powerful computing power and elastic scaling features of cloud computing to meet the needs of large-scale data processing.

[0113] Edge computing support: For some application scenarios, it is also possible to consider deploying the model on edge devices, such as embedded systems or Internet of Things (IoT) terminals, to enable real-time data analysis and local decision-making. This is suitable for dissolved gas concentration prediction tasks in a variety of similar application scenarios.

[0114] In the embodiments of the present invention, technical features not described in detail are all existing technologies or conventional technical means and will not be described in detail here.

[0115] Finally, it should be noted that the above embodiments are only specific implementation methods of the present invention, which are used to illustrate the technical solutions of the present invention, rather than to limit them, and the protection scope of the present invention is not limited thereto. Those skilled in the art should understand that any person skilled in the art can modify or easily think of changes to the technical solutions described in the above embodiments within the technical scope disclosed by the present invention, or make equivalent replacements for some of the technical features therein; and these modifications, changes or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included in the protection scope of the present invention.

Claims

1. A method for predicting dissolved gas concentration in transformer oil, which collects dissolved gas concentration data in transformer oil and generates a time series data set of dissolved gas concentration in transformer oil, characterized in that: The following steps are involved: Applying adaptive noise complete ensemble empirical mode decomposition to process the time series data set of dissolved gas concentration in the transformer oil, a number of intrinsic mode functions with different frequency characteristics are obtained; Build a hybrid deep learning model based on a temporal convolutional network, a bidirectional long short-term memory network, and a transformer architecture. The temporal convolutional network is used to capture local features, the bidirectional long short-term memory network is responsible for processing long-term dependencies, and the transformer uses a self-attention mechanism to achieve information exchange. All intrinsic mode functions are divided into a training set and a test set, the training set is input into a hybrid deep learning model for training to obtain a prediction model for dissolved gas concentration in transformer oil, and the test set is input into the prediction model for dissolved gas concentration in transformer oil for testing, verification and optimization to obtain an optimized prediction model for dissolved gas concentration in transformer oil; Finally, the optimization model for predicting dissolved gas concentration in transformer oil is used to give the final predicted value of dissolved gas concentration in transformer oil.

2. The method for predicting dissolved gas concentration in transformer oil according to claim 1, wherein: Adaptive noise complete ensemble empirical mode decomposition (EMD) is applied to process the time series dataset of dissolved gas concentration in transformer oil. The specific method is as follows: the time series dataset is denoised and smoothed. By adding white noise of different phases multiple times and then averaging the data, the data is decomposed into several physically meaningful intrinsic mode functions. Each intrinsic mode function represents the fluctuation characteristics of the original data within a specific frequency range.

3. The method for predicting dissolved gas concentration in transformer oil according to claim 2, wherein: Each intrinsic mode function is adjusted through interpolation method so that their time points correspond one to one; the prediction sequence of each intrinsic mode function is padded or trimmed. When the prediction sequence of a certain intrinsic mode function is longer than other sequences, the redundant data points are trimmed from both ends. If the prediction sequence of the intrinsic mode function is short, the last valid value is added to the end of the prediction sequence or linear interpolation is used to supplement the missing data points. After data padding or trimming, the prediction results of each intrinsic mode function have equal length within the same time range.

4. The method for predicting dissolved gas concentration in transformer oil according to claim 1, wherein: The K-fold cross-validation method is used to divide the training data into K non-overlapping subsets, which are used as test sets in turn to evaluate the prediction model.

5. The method for predicting dissolved gas concentration in transformer oil according to claim 4, wherein: Configure the optimizer and loss function for the hybrid deep learning model of each intrinsic mode function and compile the hybrid deep learning model; use the Adam optimizer as the optimizer and the mean square error as the loss function.

6. The method for predicting dissolved gas concentration in transformer oil according to claim 5, wherein: Use Python for programming, combined with the TensorFlow deep learning framework; use Docker container technology to encapsulate the learning model; deploy the hybrid deep learning model to a cloud server, or deploy the hybrid deep learning model to an edge device.

7. The method for predicting dissolved gas concentration in transformer oil according to any one of claims 1 to 6, characterized in that: The method for collecting dissolved gas concentration data in transformer oil is to regularly extract oil samples from the operating transformer and measure the dissolved gas concentration in the transformer oil using a portable or online monitoring device.

Images