Water turbine bearing temperature time sequence prediction method, system and equipment based on GRU and LSTM hybrid network and medium

Abstract

The invention discloses a water turbine bearing temperature time sequence prediction method, system and device based on a GRU and LSTM hybrid network and a medium, and belongs to the technical field of water turbine bearing temperature time sequence prediction, and the method comprises the steps: obtaining water turbine bearing temperature monitoring data, and carrying out the data preprocessing, and obtaining the preprocessed temperature data; constructing an input pair and an output pair by the preprocessed temperature data according to a set time step, and uniformly standardizing an input feature and an output target to the same numerical value interval to obtain a time sequence sample; and inputting the time sequence sample into a parallel deep learning model comprising a long short-term memory network branch and a gating circulation unit branch, and carrying out adaptive weighted fusion on output features of the two branches through an attention mechanism to obtain prediction output. The advantages of the LSTM and the GRU are fully utilized, long-term dependence and short-term dynamics are captured at the same time, the comprehensiveness of feature extraction is improved, the effect similar to integrated learning is introduced into the mixed structure, and the over-fitting risk is reduced.

Description

Technical Field

[0001] This invention relates to the field of time-series prediction technology for turbine bearing temperature, specifically to a method, system, device, and medium for time-series prediction of turbine bearing temperature based on a hybrid network of GRU and LSTM. Background Technology

[0002] As the core equipment in hydropower generation, the safe and stable operation of the turbine is crucial to the power system. Bearing temperature is one of the key indicators of turbine operating status; abnormal fluctuations often indicate potential equipment failure or performance degradation. Accurately predicting turbine bearing temperature trends not only enables timely detection of potential faults but also provides a scientific basis for preventative maintenance, effectively reducing equipment repair costs and avoiding downtime losses due to sudden failures. In recent years, with the widespread application of IoT technology and sensor equipment in hydropower stations, large amounts of temperature time-series data have been collected and stored in real time.

[0003] However, turbine bearing temperature is affected by a variety of factors, including load fluctuations, ambient temperature, and cooling system efficiency, exhibiting highly nonlinear and time-varying characteristics. Traditional prediction methods, such as time series analysis and regression models, mostly employ a single type of model structure, failing to fully consider the simultaneous long-term trends and short-term fluctuations in temperature data, and thus struggling to effectively capture these complex dependencies.

[0004] Deep learning methods, especially recurrent neural networks (RNNs) and their variants, have shown great potential in the field of time series prediction due to their ability to handle sequential data and capture temporal dependencies. Long Short-Term Memory (LSTM) networks effectively solve the gradient vanishing problem of traditional RNNs by introducing gating mechanisms and memory units, and can capture long-term dependencies better. Gated Recurrent Units (GRUs), as a simplified version of LSTMs, have a more streamlined architecture, fewer parameters, and higher computational efficiency, and perform no less well than LSTMs in certain tasks. However, single-structure LSTM or GRU still have limitations when dealing with complex time series data: LSTM finely controls long-term dependencies through input gates, forget gates, output gates, and cell states, making it suitable for handling complex time series patterns. However, when dealing with short sequence tasks, the fine gating of LSTM (such as forget gates and input gates) may learn noisy or irrelevant features, and its complex structure may lead to gradient instability. GRU has fewer parameters, faster training, and more direct gradient propagation, and performs better in some short-term memory tasks. However, its ability to model long sequences may be weaker than that of LSTM, and it is more prone to overfitting when dealing with long sequence problems. Summary of the Invention

[0005] In view of the above-mentioned problems, the present invention is proposed.

[0006] Therefore, the technical problem solved by this invention is: how to address the limitations of single-structure LSTM or GRU when processing complex time-series data. LSTM, through input gates, forget gates, output gates, and cell states, finely controls long-term dependencies and is suitable for processing complex time-series patterns. However, when processing short-series tasks, LSTM's fine gating may learn noisy or irrelevant features, and its complex structure may lead to gradient instability. GRU has fewer parameters, faster training, and more direct gradient propagation, performing better in some short-term memory tasks. However, its ability to model long sequences may be weaker than LSTM, and it is more prone to overfitting when dealing with long-series problems.

[0007] To address the aforementioned technical problems, this invention provides the following technical solution: a time-series prediction method for turbine bearing temperature based on a hybrid GRU and LSTM network, comprising: acquiring turbine bearing temperature monitoring data and performing data preprocessing to obtain preprocessed temperature data; constructing input-output pairs from the preprocessed temperature data according to a set time step, and standardizing the input features and output target to the same numerical range to obtain time-series samples; inputting the time-series samples into a parallel deep learning model including a long short-term memory network branch and a gated recurrent unit branch, and adaptively weighting and fusing the output features of the two branches through an attention mechanism to obtain the predicted output.

[0008] As a preferred embodiment of the time-series prediction method for turbine bearing temperature based on a hybrid GRU and LSTM network described in this invention, the step of acquiring turbine bearing temperature monitoring data and performing data preprocessing to obtain preprocessed temperature data includes: performing preliminary checks on the temperature data, including basic statistical information analysis and missing value detection; deleting zero-value data; calculating the median and absolute median difference based on a sliding window, determining a threshold, identifying whether the current data point is a non-zero outlier, and if it is a non-zero outlier, replacing the non-zero outlier data point with the median of the window.

[0009] As a preferred embodiment of the time-series prediction method for turbine bearing temperature based on a hybrid GRU and LSTM network described in this invention, the step of constructing input-output pairs from preprocessed temperature data according to a set time step, and standardizing the input features and output targets to the same numerical range to obtain time-series samples includes: selecting continuous temperature data as input features; selecting temperature data at corresponding future time points as output targets; and transforming the input features and output targets using the same standardization scaler to scale them to a preset numerical range.

[0010] As a preferred embodiment of the time-series prediction method for turbine bearing temperature based on a hybrid GRU and LSTM network described in this invention, the step of inputting time-series samples into a parallel deep learning model including a long short-term memory network branch and a gated recurrent unit branch, and adaptively weighting and fusing the output features of the two branches through an attention mechanism to obtain the predicted output includes: inputting time-series samples into two deep learning branches respectively, one branch including two long short-term memory network layers and the other branch including two gated recurrent unit layers; extracting the output features of the two branches respectively; calculating the learnable weight coefficients of the two branches through an attention mechanism, and weighting and fusing the output features of the two branches.

[0011] As a preferred embodiment of the turbine bearing temperature time series prediction method based on GRU and LSTM hybrid network described in this invention, the parallel deep learning model includes multiple DropOut layers; a DropOut layer is set after each long short-term memory network layer; a DropOut layer is set after each gated recurrent unit layer; and a DropOut layer is set before the fully connected layer connected before the output.

[0012] This preferred scheme effectively reduces feature redundancy and model overfitting during training by setting DropOut layers before the Long Short-Term Memory network layer, the gated recurrent unit layer, and the fully connected layer, thereby improving the model's generalization ability to test data and making the prediction curve more stable.

[0013] As a preferred embodiment of the turbine bearing temperature time series prediction method based on GRU and LSTM hybrid network described in this invention, the parallel deep learning model further includes using mean square error as a loss function; defining the loss function as the mean square error between the predicted output and the actual target; selecting the Adam optimizer; and iteratively updating the parameters of the parallel deep learning model using the loss function and the optimizer.

[0014] This preferred scheme defines mean squared error as the loss function and uses the Adam optimizer to iteratively update parameters, so that the difference between the predicted output and the actual target can be accurately measured and converged quickly, reducing the number of training rounds and improving the convergence speed and prediction accuracy of the model.

[0015] As a preferred embodiment of the turbine bearing temperature time series prediction method based on a hybrid GRU and LSTM network described in this invention, the adaptive weighted fusion of the output features of the two branches through an attention mechanism includes: extracting the output features of the Long Short-Term Memory (LSTM) network branch and the gated recurrent unit (GRU) branch; generating learnable fusion weight coefficients based on the content of the current input features; automatically increasing the fusion weight of the LSTM network branch when the input features contain patterns requiring fine-grained memory control; and automatically increasing the fusion weight of the GRU branch when the input features are simple patterns.

[0016] This preferred scheme increases the weight of the long short-term memory network branch when dealing with complex input features and increases the weight of the gated recurrent unit branch when dealing with simple input features through an attention mechanism. This achieves adaptive adjustment of branch contributions, enhances the model's adaptability to different temperature change patterns, and ensures that the prediction results maintain high accuracy in both volatile and stable scenarios.

[0017] This invention provides a time-series prediction system for turbine bearing temperature based on a hybrid network of GRU and LSTM.

[0018] To address the aforementioned technical problems, this invention provides the following technical solution: a time-series prediction system for turbine bearing temperature based on a hybrid GRU and LSTM network, comprising: a data acquisition module, a sample construction module, and a parallel modeling module; the data acquisition module is used to acquire turbine bearing temperature monitoring data and perform data preprocessing to obtain preprocessed temperature data; the sample construction module is used to construct input-output pairs from the preprocessed temperature data according to a set time step, and to standardize the input features and output targets to the same numerical range to obtain time-series samples; the parallel modeling module is used to input the time-series samples into a parallel deep learning model including a long short-term memory network branch and a gated recurrent unit branch, and to adaptively weight and fuse the output features of the two branches through an attention mechanism to obtain the predicted output.

[0019] The present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, characterized in that the processor executes the computer program to implement the steps of the above-described method for predicting the time series temperature of a water turbine bearing based on a hybrid GRU and LSTM network.

[0020] The present invention provides a computer-readable storage medium having a computer program stored thereon, characterized in that, when the computer program is executed by a processor, it implements the steps of the aforementioned method for predicting the timing of turbine bearing temperature based on a hybrid GRU and LSTM network.

[0021] The beneficial effects of this invention are as follows: This invention proposes a hybrid GRU and LSTM network specifically for the accurate prediction of turbine bearing temperature. The architecture employs a parallel dual-branch structure, simultaneously utilizing LSTM and GRU to extract features from time-series data, and then achieving adaptive fusion of the outputs of the two branches through an attention mechanism. This design allows the model to fully leverage the advantages of LSTM and GRU, while capturing both long-term dependencies and short-term dynamics, thus improving the comprehensiveness of feature extraction. Furthermore, the hybrid structure itself can introduce effects similar to ensemble learning, reducing the risk of overfitting. Attached Figure Description

[0022] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 The following is a flowchart of a method for predicting the time series temperature of a turbine bearing based on a hybrid network of GRU and LSTM, provided as an embodiment of the present invention.

[0024] Figure 2 This image shows a comparison of temperature data before and after cleaning in a time-series prediction method for turbine bearing temperature based on a hybrid GRU and LSTM network, provided as an embodiment of the present invention.

[0025] Figure 3 The diagram shows the LSTM network structure of a turbine bearing temperature time series prediction method based on a hybrid GRU and LSTM network, as provided in one embodiment of the present invention.

[0026] Figure 4 The diagram shows the GRU network structure of a turbine bearing temperature time series prediction method based on a hybrid GRU and LSTM network, as provided in one embodiment of the present invention.

[0027] Figure 5 The diagram shows the structure of a hybrid network (GRU and LSTM) for a turbine bearing temperature time-series prediction method based on a hybrid network (GRU and LSTM) provided in one embodiment of the present invention.

[0028] Figure 6 The graph shows the variation of the loss function of three models for a time-series prediction method for turbine bearing temperature based on a hybrid network of GRU and LSTM, provided as an embodiment of the present invention.

[0029] Figure 7The image shows the prediction effect of the GRU model for a time-series prediction method of turbine bearing temperature based on a hybrid network of GRU and LSTM, provided as an embodiment of the present invention.

[0030] Figure 8 The image shows the prediction effect of an LSTM model for a time-series prediction method for turbine bearing temperature based on a hybrid network of GRU and LSTM, provided as an embodiment of the present invention.

[0031] Figure 9 The image shows the prediction effect of an LSTM-GRU hybrid model for a time-series prediction method of turbine bearing temperature based on a hybrid network of GRU and LSTM, provided as an embodiment of the present invention.

[0032] Figure 10 This is a comparison chart of three models for evaluating a time-series prediction method for turbine bearing temperature based on a hybrid GRU and LSTM network, provided as an embodiment of the present invention. Detailed Implementation

[0033] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0034] Example 1, referring to Figure 1 This is one embodiment of the present invention, which provides a method for predicting the time series temperature of turbine bearings based on a hybrid GRU and LSTM network, including:

[0035] S1. Obtain the temperature monitoring data of the turbine bearing and perform data preprocessing to obtain the preprocessed temperature data.

[0036] S2. Construct input-output pairs from the preprocessed temperature data according to the set time steps, and standardize the input features and output targets to the same numerical range to obtain time series samples.

[0037] S3. Input the time-series samples into a parallel deep learning model that includes a long short-term memory network branch and a gated recurrent unit branch. Adaptively weight and fuse the output features of the two branches through an attention mechanism to obtain the predicted output.

[0038] It should be noted that during long-term operation, the bearing temperature of a hydro turbine is affected by multiple factors such as hydraulic load fluctuations, mechanical friction, and the operating environment, exhibiting a clear temporal variation characteristic. Without effective prediction and early warning, abnormal temperature rises may lead to decreased equipment operating efficiency or even equipment failure and shutdown.

[0039] Therefore, through steps S1-S3, a deep learning model is used to model the bearing temperature. First, the original monitoring data is cleaned and standardized to ensure the validity and consistency of the input data. Second, through the parallel structure of LSTM and GRU branches, the long-term dependency features and short-term variation features of complex time series data are fully extracted. Finally, by combining the adaptive weighted fusion of the attention mechanism, the contribution ratio of the two branches under different data modes is dynamically balanced, thereby achieving accurate prediction of the turbine bearing temperature.

[0040] Example 2, refer to Figures 2-5 As an embodiment of the present invention, based on the previous embodiment, a method for predicting the time series temperature of a turbine bearing based on a hybrid network of GRU and LSTM is provided, comprising:

[0041] In the embodiments of this application, data preprocessing can perform preliminary checks on temperature data, including basic statistical information analysis and missing value detection, deletion of zero-value data, calculation of median and absolute median difference based on sliding window, determination of threshold, identification of non-zero outliers, and replacement of outliers with the median of the window.

[0042] In an alternative implementation, data preprocessing may also employ a Z-score-based method to detect outliers in the temperature data and replace data points that deviate from the mean by more than a set threshold.

[0043] In another alternative implementation, data preprocessing can also employ a wavelet transform-based filtering method to denoise the temperature data and eliminate high-frequency noise components.

[0044] This invention, through data preprocessing, can effectively remove zero values ​​and outliers, correct abrupt changes in temperature curves, eliminate noise interference, and ensure the authenticity and smoothness of input data, thereby improving the convergence speed and prediction accuracy of the prediction model.

[0045] Furthermore, in step S1, the turbine bearing temperature monitoring data is acquired and preprocessed to obtain preprocessed temperature data, including the following steps A1-A3:

[0046] A1. Conduct a preliminary check on the temperature data, including basic statistical information analysis and missing value detection;

[0047] A2. Delete zero-value data;

[0048] A3. Calculate the median and absolute median difference based on the sliding window, determine the threshold, identify whether the current data point is a non-zero outlier, and replace the non-zero outlier data point with the median in the window if it is a non-zero outlier.

[0049] In this embodiment of the application, the threshold determination in step A3 can be based on the calculation of the median and the absolute median difference using a sliding window, constructing a judgment condition to identify whether the current data point deviates from the median by more than a set range. If the condition is met, the data point is marked as an outlier.

[0050] In one alternative implementation, the threshold can also be determined using a statistical method of mean and standard deviation, identifying data points that exceed the mean by several times the standard deviation as outliers.

[0051] In another alternative implementation, the threshold can also be determined using a quantile-based method, by setting an upper quantile and a lower quantile interval, and identifying data points falling outside the interval as outliers.

[0052] This invention identifies and corrects outliers by determining a threshold, which can effectively eliminate data points that deviate from normal variation patterns, reduce noise interference, ensure the authenticity and continuity of input data, and provide a more reliable data foundation for prediction models.

[0053] Specifically, turbine bearing temperature data typically contains various noises and outliers, which may originate from sensor malfunctions, measurement errors, or recording mistakes. To ensure that the constructed prediction model can predict based on actual temperature changes rather than noise, a preliminary check of the data is first performed, including basic statistical information analysis and missing value detection. Statistical analysis shows that there are a small number of zero values ​​and outliers in the raw data, which can affect the model's training performance. For zero values, a direct removal strategy is adopted because in bearing temperature monitoring, zero values ​​usually represent sensor malfunctions or data recording interruptions and have no actual physical meaning.

[0054] For the detection and handling of non-zero outliers, a Hampel filter is used. The Hampel filter works by calculating the median and median absolute deviation (MAD) through a sliding window, identifying data points that deviate from the median within the window by more than a preset threshold as outliers. The specific process is as follows:

[0055] Set a sliding window; calculate the median and absolute median difference (MAD) of the data within the window; set a threshold of 3 times MAD multiplied by an adjustment factor of 1.4826 to make MAD comparable to the standard deviation of the Gaussian distribution; if the absolute difference between the current point and the median of the window exceeds the threshold, mark it as an outlier and replace it with the median of the window.

[0056] After processing with the Hampel filter, multiple outliers in the original data were successfully identified and corrected, effectively smoothing abrupt changes and noise in the temperature curve while preserving the true temperature change trend. For example... Figure 2The image shows a comparison of the data before and after cleaning. It can be seen that the temperature curve after cleaning is smoother and more consistent, and better reflects the physical characteristics of bearing temperature changes.

[0057] In this embodiment, the input features and output targets are uniformly standardized to the same numerical range. The time series samples obtained can be transformed by the same normalization scaler to scale them to the [0,1] range.

[0058] In one alternative implementation, the input features and output target are standardized to the same numerical range. The time series samples can also be obtained by using the Z-score standardization method to convert the input features and output target into values ​​with a mean of zero and a variance of one, respectively.

[0059] In another alternative implementation, the input features and output targets are standardized to the same numerical range. The time series samples can also be obtained by using the maximum absolute value scaling method to simultaneously scale the input features and output targets to the range based on the maximum absolute value.

[0060] This invention ensures that the proportional relationship between input features and output targets is not disrupted through unified standardization, avoids the impact of numerical range differences on the training process, enables the model to learn efficiently in the same numerical space, and improves the stability and convergence speed of prediction.

[0061] Furthermore, in step S2, the preprocessed temperature data is used to construct input-output pairs according to a set time step, and the input features and output targets are standardized to the same numerical range to obtain time-series samples, including the following steps B1-B3:

[0062] B1. Select continuous temperature data as input features.

[0063] B2. Select the temperature data at the corresponding future time point as the output target.

[0064] B3. Apply the same standardized scaler to both the input features and the output target to scale them to a preset numerical range.

[0065] Specifically, after outlier handling, a sliding window method is used to create input-output pairs to construct time-series samples suitable for deep learning model training. Specifically, temperature data from the past 50 time steps are selected as input features, and the temperature value from the 10th future time step is used as the output target. Simultaneously, to accelerate model convergence and improve numerical stability, standardization is applied to all features and target values, scaling the data to the [0,1] interval. The same scaler is used for both input features and output targets during standardization to maintain their proportional relationship. Finally, the preprocessed data is divided into training and test sets in an 8:2 ratio, with the training set used for model training and parameter optimization, and the test set used to evaluate the model's generalization performance.

[0066] Furthermore, in step S3, the temporal samples are input into a parallel deep learning model that includes a long short-term memory network branch and a gated recurrent unit branch. The output features of the two branches are adaptively weighted and fused through an attention mechanism to obtain the predicted output, including the following steps C1-C3:

[0067] C1. Input the time series samples into two deep learning branches respectively. One branch includes two long short-term memory network layers, and the other branch includes two gated recurrent unit layers.

[0068] C2. Extract the output features of the two branches respectively.

[0069] C3. Calculate the learnable weight coefficients of the two branches through the attention mechanism, and perform weighted fusion of the output features of the two branches.

[0070] Specifically, in step C1, LSTM is a classic deep learning model for processing sequential data, specifically designed to solve the gradient vanishing problem in RNNs. LSTM effectively captures long-term dependencies by introducing memory cells and three gating mechanisms: input gate, forget gate, and output gate. Its core advantage lies in its ability to remember long-term information in the sequence while forgetting irrelevant information, making it particularly suitable for processing time-series data with long-term trends and periodic changes, such as the temperature of turbine bearings. This invention designs an LSTM network with 2 LSTM layers, 1 Dense layer, 3 DropOut layers, 1 input layer, and 1 output layer. The network structure is as follows: Figure 3 As shown, the model uses the Adam optimizer with an initial learning rate of 0.001 and a loss function of Mean Squared Error (MSE), which directly measures the difference between the predicted and true values. Furthermore, an early stopping strategy is implemented during training, with the patience parameter set to 10, meaning that if the model does not achieve better results after 10 iterations, the iteration process stops.

[0071] GRU is a simplified variant of LSTM. By merging the forget gate and input gate in LSTM into an update gate and introducing a reset gate, it significantly reduces the number of model parameters while maintaining the ability to capture long-term dependencies. Compared to LSTM, GRU has a more streamlined structure and higher computational efficiency, and in some tasks, it can even achieve performance comparable to or better than LSTM. For better horizontal comparison, this invention also designs an LSTM network with 2 GRU layers, 1 Dense layer, 3 DropOut layers, 1 input layer, and 1 output layer. The network structure is as follows: Figure 4 As shown in the figure. The model also uses the Adam optimizer, with an initial learning rate set to 0.001 and the loss function set to MSE. An early stopping strategy was also added during training.

[0072] In this embodiment of the application, the learnable weight coefficients of the two branches in step C3 can be calculated by using the attention mechanism based on the input temporal sample features. The dynamic weight coefficients corresponding to the two branches can be generated by the attention mechanism, and the output features of the long short-term memory network branch and the gated recurrent unit branch can be weighted and fused according to the weight coefficients.

[0073] In an alternative implementation, the learnable weight coefficients of the two branches can be calculated using the dot product attention method, which calculates the weights based on the similarity between the input features and the branch outputs.

[0074] In another alternative implementation, the learnable weight coefficients of the two branches can be calculated using an additive attention method, which generates weights by processing the input features through a trainable feedforward network.

[0075] This invention dynamically calculates and allocates the weights of the two branches through an attention mechanism, thereby increasing the contribution of the long short-term memory network branch when processing complex features and increasing the contribution of the gated recurrent unit branch when processing simple features, thus improving the model's adaptability and prediction accuracy under different input modes.

[0076] Furthermore, the parallel deep learning model includes multiple DropOut layers; a DropOut layer is set after each Long Short-Term Memory network layer; a DropOut layer is set after each gated recurrent unit layer; and a DropOut layer is set before the fully connected layer connected before the output.

[0077] Furthermore, the parallel deep learning model also includes using mean squared error as the loss function; defining the loss function as the mean squared error between the predicted output and the actual target; selecting the Adam optimizer; and iteratively updating the parameters of the parallel deep learning model using the loss function and the optimizer.

[0078] Furthermore, step C3 involves adaptively weighting and fusing the output features of the two branches using an attention mechanism, including the following steps C31-C34:

[0079] C31. Extract the output features of the branches of the Long Short-Term Memory network and the gated recurrent unit branches;

[0080] C32. Generate learnable fusion weight coefficients based on the content of the current input features;

[0081] C33. When the input features contain patterns that require fine-grained memory control, automatically increase the fusion weights of the Long Short-Term Memory network branches;

[0082] C34. When the input features are simple structural patterns, automatically increase the fusion weight of the gated loop unit branches.

[0083] Specifically, the LSTM-GRU hybrid model proposed in this invention is an innovative deep learning architecture designed to combine the advantages of LSTM and GRU while overcoming their limitations. This model employs a parallel dual-branch structure, extracting temporal features through LSTM and GRU respectively, and then achieving adaptive fusion through an attention mechanism. Figure 5 The diagram shows the designed hybrid network structure. The key innovation of the hybrid model lies in its attention-weighted fusion mechanism. This mechanism does not simply connect or average the outputs of the two branches, but dynamically adjusts the contribution ratio of the LSTM and GRU branches through learnable weight coefficients. When the input data contains complex patterns requiring fine-grained memory control, the model increases the weight of the LSTM branch; when dealing with relatively simple patterns, the model tends to assign greater weight to the computationally more efficient GRU branch. This adaptive fusion strategy allows the hybrid model to flexibly handle different types of features in the bearing temperature sequence. Similarly, the model uses the Adam optimizer with an initial learning rate of 0.001 and the MSE loss function, and also incorporates an early stopping strategy during training.

[0084] Example 3, referring to Figures 6-10 As one embodiment of the present invention, a method for predicting the time series temperature of a turbine bearing based on a hybrid network of GRU and LSTM is provided. To verify the beneficial effects of the present invention, scientific demonstration is carried out through experiments.

[0085] The data has already been preprocessed during the data preprocessing stage to ensure a high-quality dataset. During training, the LSTM, GRU, and LSTM-GRU hybrid models were all trained with 30 iterations (epochs) and a batch size of 32. All models dynamically partitioned the validation set to 20% using the validation_split parameter. MSE was used as the loss function for all models during training. The changes in the loss function during training for the three models are shown below. Figure 6 As shown in the figure, both the LSTM model and the hybrid model exhibit good convergence in both training and validation losses, with the hybrid model converging to a lower level. However, the validation loss of the GRU model shows a pattern of first decreasing and then increasing, indicating a significant overfitting issue during training.

[0086] Next, the last 300 data points were selected from the test set to verify the prediction performance of the three trained models. The fitting performance graphs of the three models are shown below. Figure 7 As shown in 8 and 9.

[0087] Considering the fitting performance of the three models, the GRU model's fitting effect is clearly inferior to the other two models' prediction performance, which confirms the overfitting phenomenon that occurred during training. The LSTM model, on the other hand, shows a smoother fit to data points 50 to 300 compared to the mixed model, indicating that the LSTM model focuses more on long-term trends and is slightly less adept at capturing short-term trends.

[0088] Next, we will compare the performance of the three models in terms of four aspects: MSE, Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), and Coefficient of Determination (R²) score. Figure 10 As shown in the figures, the experimental results clearly demonstrate that the LSTM-GRU hybrid model exhibits significant performance advantages in the turbine bearing temperature prediction task. Looking at the four key evaluation metrics, the hybrid model's MSE is 0.2144, RMSE is 0.4630, and MAE is 0.3113, all significantly lower than the GRU model. In terms of R², the hybrid model achieves a high value of 0.9825, slightly higher than the LSTM model's 0.9820, and far superior to the GRU model's 0.7025. These experimental data strongly demonstrate the effectiveness of the attention fusion mechanism proposed in this invention in adaptively integrating the advantages of LSTM and GRU.

[0089] Example 4 is an embodiment of the present invention. This embodiment provides a time series prediction system for turbine bearing temperature based on a hybrid network of GRU and LSTM, including a data acquisition module, a sample construction module and a parallel modeling module.

[0090] The data acquisition module is used to acquire temperature monitoring data of the turbine bearings and perform data preprocessing to obtain preprocessed temperature data.

[0091] The sample construction module is used to construct input-output pairs from the preprocessed temperature data according to a set time step, and to standardize the input features and output targets to the same numerical range to obtain time series samples.

[0092] The parallel modeling module is used to input time-series samples into a parallel deep learning model that includes a long short-term memory network branch and a gated recurrent unit branch. The output features of the two branches are adaptively weighted and fused through an attention mechanism to obtain the predicted output.

[0093] This embodiment also provides an electronic device applicable to a turbine bearing temperature timing prediction method based on a hybrid GRU and LSTM network, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the turbine bearing temperature timing prediction method based on a hybrid GRU and LSTM network as proposed in the above embodiment.

[0094] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements a turbine bearing temperature timing prediction method based on a hybrid GRU and LSTM network as proposed in the above embodiment.

[0095] The storage medium proposed in this embodiment belongs to the same inventive concept as the method for predicting the timing of turbine bearing temperature based on a hybrid GRU and LSTM network proposed in the above embodiments. Technical details not described in detail in this embodiment can be found in the above embodiments, and this embodiment has the same beneficial effects as the above embodiments.

[0096] Based on the above description of the implementation methods, those skilled in the art can clearly understand that the present invention can be implemented using software and necessary general-purpose hardware, and of course, it can also be implemented using hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk, or optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of the various embodiments of the present invention.

[0097] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for predicting the time series temperature of turbine bearings based on a hybrid GRU and LSTM network, characterized in that: include, Acquire temperature monitoring data of the turbine bearings and perform data preprocessing to obtain preprocessed temperature data; The preprocessed temperature data is used to construct input-output pairs according to a set time step, and the input features and output targets are standardized to the same numerical range to obtain time series samples. The time-series samples are input into a parallel deep learning model that includes a long short-term memory network branch and a gated recurrent unit branch. The output features of the two branches are adaptively weighted and fused through an attention mechanism to obtain the predicted output.

2. The method for predicting the time series temperature of turbine bearings based on a hybrid GRU and LSTM network as described in claim 1, characterized in that: The process of acquiring turbine bearing temperature monitoring data and performing data preprocessing to obtain preprocessed temperature data includes, A preliminary check of the temperature data was performed, including basic statistical information analysis and missing value detection. Delete zero-value data; The median and absolute median difference are calculated based on a sliding window. A threshold is determined, and the current data point is identified as a non-zero outlier. If it is a non-zero outlier, the non-zero outlier data point is replaced with the median in the window.

3. The method for predicting the time series temperature of turbine bearings based on a hybrid GRU and LSTM network as described in claim 2, characterized in that: The process involves constructing input-output pairs from the preprocessed temperature data according to a set time step, and standardizing the input features and output targets to the same numerical range to obtain time-series samples. Continuous temperature data is selected as the input feature. Select the temperature data at the corresponding future time point as the output target; The input features and output target are transformed using the same normalized scaler to scale them to a preset numerical range.

4. The method for predicting the time series temperature of turbine bearings based on a hybrid GRU and LSTM network as described in claim 3, characterized in that: The process involves inputting time-series samples into a parallel deep learning model that includes branches of Long Short-Term Memory (LSTM) and Gated Recurrent Units (GRU). An attention mechanism is used to adaptively weight and fuse the output features of the two branches to obtain a predicted output, including... The time series samples are input into two deep learning branches, one of which includes two long short-term memory network layers and the other includes two gated recurrent unit layers. Extract the output features of the two branches separately; The learnable weight coefficients of the two branches are calculated using an attention mechanism, and the output features of the two branches are then weighted and fused.

5. The method for predicting the time series temperature of turbine bearings based on a hybrid GRU and LSTM network as described in claim 4, characterized in that: The parallel deep learning model includes multiple DropOut layers; Set a DropOut layer after each Long Short-Term Memory (LSTM) network layer; Set a DropOut layer after each gated loop unit layer; Set a DropOut layer before the fully connected layer connected before the output.

6. The method for predicting the time series temperature of turbine bearings based on a hybrid GRU and LSTM network as described in claim 5, characterized in that: The parallel deep learning model also includes the use of mean squared error as a loss function; The loss function is defined as the mean squared error between the predicted output and the actual target. Select the Adam optimizer; The parameters of the parallel deep learning model are iteratively updated using a loss function and an optimizer.

7. The method for predicting the time series temperature of turbine bearings based on a hybrid GRU and LSTM network as described in claim 6, characterized in that: The adaptive weighted fusion of the two-branch output features through an attention mechanism includes, Extract the output features of branches of the Long Short-Term Memory network and gated recurrent unit branches; Generate learnable fusion weight coefficients based on the content of the current input features; When the input features contain patterns that require fine-grained memory control, the fusion weights of the long short-term memory network branches are automatically increased; When the input features are simple structural patterns, the fusion weight of the gated recurrent unit branches is automatically increased.

8. A time-series prediction system for turbine bearing temperature based on a hybrid GRU and LSTM network, employing the time-series prediction method for turbine bearing temperature based on a hybrid GRU and LSTM network as described in any one of claims 1 to 7, characterized in that, include: Data acquisition module, sample construction module, and parallel modeling module; The data acquisition module is used to acquire the temperature monitoring data of the turbine bearing and perform data preprocessing to obtain the preprocessed temperature data. The sample construction module is used to construct input-output pairs from the preprocessed temperature data according to a set time step, and to standardize the input features and output targets to the same numerical range to obtain time-series samples. The parallel modeling module is used to input time-series samples into a parallel deep learning model that includes a long short-term memory network branch and a gated recurrent unit branch. The output features of the two branches are adaptively weighted and fused through an attention mechanism to obtain the predicted output.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method for predicting the timing of turbine bearing temperature based on a hybrid GRU and LSTM network as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the time-series prediction method for turbine bearing temperature based on a hybrid network of GRU and LSTM as described in any one of claims 1 to 7.

Images