Methods and systems for predicting the performance degradation trend of hydropower units considering operating parameters

Abstract

This invention discloses a method and system for predicting the performance degradation trend of hydropower units considering operating parameters, belonging to the field of hydropower unit condition assessment and early warning. This invention employs a probability interval prediction model, which consists of a multi-head attention module, a Bi-GRU module, and a quantile loss module. The input includes not only the historical performance degradation degree of the unit but also the unit's operating parameters. First, the multi-head attention module is used to extract hidden correlation features between the unit's operating parameters; second, the Bi-GRU module is used to extract the hidden feature map of the operating conditions and the temporal features of the unit's degradation degree; finally, the quantile loss module obtains the confidence probability prediction interval result of the unit's performance degradation degree. This overcomes the shortcomings of inaccurate prediction of unit performance degradation caused by abnormal unit vibration and sway due to environmental factors, achieving accurate prediction of the unit's performance degradation trend and improving the reliability of the model prediction. Because it considers the unit's operating parameter information, it has stronger robustness.

Description

Technical Field

[0001] This invention belongs to the field of hydropower unit condition assessment and early warning, and more specifically, relates to a method and system for predicting the performance degradation trend of hydropower units considering operating parameters. Background Technology

[0002] Hydropower units are the core equipment in hydropower station construction. Ensuring their stable, safe, and reliable operation is a crucial task for power plant management and maintenance personnel, as it relates to the safety of the power station itself and the grid's ability to reliably provide power to society. Hydropower units are affected by multiple factors, including electromagnetic, mechanical, and hydraulic factors. With increasing cumulative operating time, abnormal vibrations can easily occur in the unit equipment, leading to fatigue, damage, and even deterioration. This affects the safe and reliable operation of the unit, further impacting the safety and stability of the power grid and causing significant socio-economic losses. Developing a theory for the safe operation and control of hydropower units under the multi-coupling constraints of water, machinery, electricity, and magnetism is an important basis for ensuring the safety of water conservancy projects. It is of great significance for improving water resource utilization, ensuring ecological water use downstream of hydropower stations, and promoting national economic development. Hydropower units undergo a process from healthy state to performance degradation, equipment failure, and even malfunction during long-term operation. Monitoring the unit's operating status during this degradation process, assessing changes in its health status, and predicting future degradation trends will enable timely detection of abnormalities, the development of reasonable maintenance plans, and the realization of condition-based maintenance for hydropower units, thus maximizing the production efficiency of power generation companies. As a crucial aspect of unit maintenance, health status assessment and performance degradation trend prediction for hydropower units have significant research and application value for improving the stable operation and maintenance of hydropower units and reducing downtime losses due to foreseeable failures.

[0003] Therefore, to ensure the safe and stable operation of hydropower units, it is urgent to conduct research on predicting the performance degradation trend of hydropower units. Based on the massive condition monitoring data of hydropower units, this study analyzes the high-dimensional function mapping relationship between unit operating parameters and unit operating status, monitors the health status of hydropower unit equipment in real time, evaluates the health performance of hydropower units during historical operation, and predicts and assesses the future health status of units. This allows for the timely detection of abnormal conditions, scientific and rational judgment of unit faults, and the arrangement of corresponding maintenance and proactive upkeep, providing decision-making technical support for fault detection and proactive maintenance, thereby ensuring the safe and stable operation of hydropower units.

[0004] Existing methods for predicting the deterioration trend of hydropower units based on vibration signals suffer from problems such as selecting overly localized and singular health indicators, making it difficult for the predicted deterioration trend results to reflect the overall operating status of the unit. Under complex operating conditions, the accuracy of the deterioration trend prediction model is insufficient, and it cannot accurately obtain the deterioration trend at future moments. In addition, the deterioration prediction model ignores the influence of historical operating condition parameters on the deterioration trend prediction, further limiting the model's predictive effectiveness and robustness. Summary of the Invention

[0005] In view of the shortcomings of the prior art, the purpose of this invention is to provide a method and system for predicting the performance degradation trend of hydropower units that takes into account operating parameters, in order to solve the problems that the prediction method for hydropower unit degradation trend is difficult to reflect the overall operating status of the unit, and has insufficient prediction accuracy and limited robustness.

[0006] To achieve the above objectives, in a first aspect, the present invention provides a method for predicting the performance degradation trend of hydropower units considering operating parameters, the method comprising:

[0007] S1. Obtain historical operating parameters of the hydropower unit and calculate the time series of degradation trends that can reflect the overall health status and deterioration performance of the hydropower unit;

[0008] S2. Input the historical operating parameters and historical deterioration trend time series of the hydropower unit into the trained probability interval prediction model to obtain the reliability prediction interval of the future deterioration of the hydropower unit.

[0009] The probability interval prediction model consists of a multi-head attention module, a Bi-GRU module, and a quantile loss module. The multi-head attention module extracts hidden feature maps of hydropower unit operating conditions from historical operating parameters of the hydropower unit. After concatenating the hidden feature maps with the historical degradation residuals of the hydropower unit, the data is input to the Bi-GRU module. The Bi-GRU module obtains the time-series features of hydropower unit degradation depth through forward GRU and backward GRU calculations, and outputs them to the quantile loss module. The quantile loss module inputs the time-series features of hydropower unit degradation depth into a feedforward fully connected layer to calculate the probability prediction results of hydropower unit degradation degree at different quantiles.

[0010] It should be noted that the samples used to train the probability interval prediction model are operating parameters and historical degradation trend time series, labeled as trend time series for future moments.

[0011] Preferably, step S1 includes:

[0012] S11. Input the operating parameters of the hydropower unit during operation into the trained single-channel health state model to obtain the oscillation health state value of each channel of the hydropower unit under the operating conditions at each time. Compare it with the real-time oscillation state parameters of the hydropower unit to obtain the degree of degradation under each channel. The single-channel health state model includes random forest and RFECV. The random forest is used to fit the relationship between the operating parameters and the oscillation state parameters. The RFECV is used to remove non-important operating parameters during the training of random forest.

[0013] S12. The health degradation of individual components during each steady-state operation of the hydropower unit is fused over time to obtain the health degradation of each component that can describe the entire steady-state process of the hydropower unit.

[0014] S13. Spatially fuse the health degradation of each component of the hydropower unit to obtain the fused degradation of the hydropower unit, thereby forming a time series of hydropower unit performance degradation trend.

[0015] It should be noted that this invention selects the oscillation of hydropower units as the research object, and uses a recursive feature elimination method based on random forest to select key operating parameters for each oscillation degree, constructing a single-channel health performance index model. This establishes the functional relationship between the hydropower unit's operating parameters and dependent variables such as oscillation degree, obtaining the health performance trends of multiple components of the hydropower unit. Based on this, a spatiotemporal model of the hydropower unit's health performance under steady-state processes is constructed, forming a model representing the performance degradation trend of a key component and the entire unit, thereby accurately reflecting the overall health performance status of the unit and obtaining the time series of the unit's performance degradation trend, providing data support for subsequent prediction tasks. The samples used to train the single-channel health state model are operating parameters, labeled as actual oscillation state parameters.

[0016] Preferably, in step S11, the degradation degree DC under each channel is... i，t The calculation formula is as follows:

[0017]

[0018] Among them, DC i，t ReLU represents the degradation degree of the hydropower unit at time t under channel i, and ReLU represents the linear rectification function. i，t V represents the actual oscillation state parameters of channel i at time t of the hydropower unit. i，t This represents the vibration and oscillation health status value of channel i of the hydropower unit under operating conditions at time t.

[0019] It should be noted that the above-mentioned degradation degree calculation formula is preferred in this invention, and ReLU can ensure that the degradation degree of a single sensing channel of the hydropower unit is not less than 0.

[0020] Preferably, in step S12, the formulas for calculating the health deterioration of each component that can describe the entire steady-state process of the hydropower unit are as follows:

[0021]

[0022]

[0023] Where time t represents the start time of a certain steady-state operation process, and D j (t) represents the degree of health degradation of component j at time t of the hydropower unit, T represents the steady-state operating time of the hydropower unit from start-up to shutdown, and D j，t This indicates Euclidean distance.

[0024] It should be noted that the present invention preferably adopts the above-mentioned time fusion method. Since the degradation degree calculation takes into account each complete steady-state operation process, the degradation degree of each component can more objectively describe the unit operation status of the entire steady-state process.

[0025] Preferably, in step S13, the formula for calculating the time series of the performance degradation trend of the spatially integrated hydropower unit is as follows:

[0026]

[0027] Where n represents the number of components in the hydropower unit, time t represents the start time of a certain steady-state operation process, and D(t) represents the time series of the hydropower unit's performance degradation trend. j (t) represents the degree of health degradation of component j of the hydropower unit at time t.

[0028] It should be noted that the present invention preferably adopts the above-mentioned spatial fusion method. Since the degradation status of each component is comprehensively considered, the obtained degradation degree can accurately reflect the overall health performance status of the unit.

[0029] Preferably, the multi-attention module includes a position encoding unit, a first residual connection unit, a masked multi-head self-attention unit, a global average pooling layer, a feedforward layer, and a second residual connection unit;

[0030] The location encoding unit is used to encode the historical operating parameters of the hydropower unit.

[0031] The first residual connection unit is used to input the residual connection position encoding and operating parameters to the masked multi-head self-attention unit;

[0032] The masked multi-head self-attention unit is used to extract multi-head attention features and input the extracted attention vector to the feedforward layer.

[0033] The feedforward layer is used to connect and merge in series to form a hidden feature map of the hydropower unit's operating conditions;

[0034] The second residual connection unit is used to connect the hidden feature map and the historical deterioration trend time series of the hydropower unit, and then input them into the Bi-GRU module.

[0035] It should be noted that the present invention preferably uses the multi-head attention module with the above-mentioned structure. It uses position encoding and multi-head attention mechanism to extract hidden correlation features between unit operating condition parameters while reducing model training time. The resulting hidden feature map of operating conditions can overcome the problem of inaccurate degradation prediction caused by environmental factors and improve the accuracy of degradation trend prediction.

[0036] To achieve the above objectives, in a second aspect, the present invention provides a hydropower unit performance degradation trend prediction system considering operating parameters, comprising: a processor and a memory;

[0037] The memory is used to store computer programs or instructions;

[0038] The processor is used to execute the computer program or instructions in the memory, causing the method described in the first aspect to be performed.

[0039] In summary, the technical solutions conceived by this invention have the following beneficial effects compared with the prior art:

[0040] This invention proposes a method for predicting the performance degradation trend of hydropower units considering operating parameters. It employs a probability interval prediction model, which consists of a multi-head attention module, a Bi-GRU module, and a quantile loss module. The model input includes not only the historical performance degradation degree of the unit but also the unit's operating parameters. First, the multi-head attention module extracts hidden correlation features among the unit's operating parameters. Second, the Bi-GRU module extracts the hidden feature map of operating conditions and the temporal features of unit degradation degree. Finally, the quantile loss module obtains the confidence probability prediction interval of the unit's performance degradation degree. This method overcomes the shortcomings of inaccurate prediction of unit performance degradation caused by abnormal vibration and sway due to environmental factors, achieving accurate prediction of unit performance degradation trends and improving the reliability of the model prediction. By considering the unit's operating parameter information, the hydropower unit performance degradation trend prediction model exhibits stronger robustness. Attached Figure Description

[0041] Figure 1 This is a flowchart of the hydropower unit deterioration trend prediction method provided in the embodiments of the present invention;

[0042] Figure 2 This is a schematic diagram of the spatiotemporal fusion degradation trend sequence results provided in an embodiment of the present invention;

[0043] Figure 3This is a structural diagram of the degradation trend prediction model provided in the embodiments of the present invention;

[0044] Figure 4 This is a schematic diagram of the degradation trend range prediction results provided in an embodiment of the present invention;

[0045] Figure 5 This is the QQ chart for predicting the degradation trend range provided in this embodiment of the invention. Detailed Implementation

[0046] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0047] This invention provides a method for predicting the degradation trend of hydropower units. The overall scheme is as follows: First, multiple single-channel health index models are established using RFECV (Recursive Feature Elimination Cross-validation) + RF (Random Forests) models. During training, the model adaptively implements the importance scoring of input features to select the optimal operating condition features. The health model finally fits the relationship between the optimal operating condition features and the vibration swing of each channel. By analyzing the difference between the vibration swing health state value and the actual swing value, the degradation index of each channel is calculated. The Euclidean distance of each channel is calculated to obtain the degradation degree of each component. The degradation degree of each component is fused over time during steady-state operation, and finally, the degradation degree of each component is fused spatially to obtain the overall degradation degree of the unit. A hydropower unit degradation trend prediction model based on Healthformer is established, and features are extracted from historical operating condition information and historical degradation degree of unit performance. Finally, the reliability prediction range of future degradation degree is predicted.

[0048] Figure 1 This is a flowchart of the hydropower unit deterioration trend prediction method provided in an embodiment of the present invention. Figure 1 As shown, this invention provides a method for predicting the performance degradation trend of hydropower units considering operating parameters, comprising the following steps:

[0049] Step (1): Using the operating parameters of the hydropower unit as the independent variables of the health state model, and the vibration and swing state parameters of the unit under health state as the dependent variables of the model, n single-channel health state index models based on RFECV+RF are established respectively. The model can be expressed as f i :X t →V i，t , i∈[1,...,n], where X t V represents the operating parameters of the hydropower unit at time t.i，t f represents the i-th dimension vibration swing measurement point at time t. i This indicates the RFECV+RF single-channel fitting model that needs to be solved in this invention.

[0050] Step (1) specifically includes the following steps:

[0051] (1-1): First, the condition monitoring data collected during the early operation phase of the hydropower unit (i.e., when the unit is still in a fully healthy state) is selected as the model dataset Z, and 80% of the samples in the dataset are randomly selected as the model training set Z. T The remaining 20% ​​is used as the model validation set Z. V .

[0052] (1-2): A random forest model was selected as the learner for feature selection, and the minimum number of feature selections, the number of features removed in each stage, and the number of cross-validations were set. The coefficient of determination R was also selected. 2 The coefficient of determination (R²) serves as an indicator for each feature importance score. 2 The calculation formula is as follows:

[0053]

[0054]

[0055] Among them, y i This represents the measured value. This represents the model's predicted value, and n represents the number of samples.

[0056] (1-3): The training process of the RFECV+RF health model can be divided into the following steps:

[0057] (1-3-1): The training dataset Z T The input is fed into the RFECV+RF model to model the original feature set, calculate the importance score of each feature, remove some features with low importance scores, update the feature set, and obtain a feature subset;

[0058] (1-3-2): Based on the importance of features determined in each stage, select different numbers of features in sequence to divide the training subset into a sub-training set and a sub-validation set. By performing k-fold cross-validation based on a random forest model on the selected feature set, determine the number of features with the highest average score.

[0059] (1-3-3): If the number of features meets the optimal feature quantity condition, then end the training, complete the feature selection of working condition parameters, and save the random forest model; otherwise, return to (1-3-1) until the optimal feature condition is met.

[0060] Step (2): Input the operating parameters of the future operation process into multiple single-channel health status models respectively, and compare the vibration health status value with the real-time vibration status value of the hydropower unit to obtain the degradation degree of each channel; calculate the Euclidean distance of the corresponding degradation degree according to the spatial position relationship of all sensing channels of each component to achieve spatial fusion; perform time fusion of the health degradation degree of single components in the steady-state operation of the hydropower unit to obtain the health degradation degree of each component that can describe the entire steady-state process of the unit; fuse the health degradation degrees of multiple components to obtain the unit fused degradation degree; specifically including the following steps:

[0061] (2-1): The operating parameters X during the future operation of the hydropower unit t′ The values ​​of vibration and sway health status V of each channel of the hydropower unit under the operating conditions at time t′ are calculated and input into the single-channel health status model. i，t′ The real-time oscillation state parameter R of the hydropower unit i，t′ By comparison, the degradation degree DC under channel i is obtained. i，t′ The calculation formula is as follows:

[0062]

[0063]

[0064] ReLU stands for Rectified Linear Activation Function, a commonly used activation function in neural networks.

[0065] (2-2): For each component, calculate its Euclidean distance based on all sensing channels of each component (for oscillation components, this includes X and Y directions; for vibration components, it includes X, Y, and Z directions):

[0066]

[0067] (2-3): Then, the health degradation of individual components during the steady-state operation of the hydropower unit is fused over time to obtain the health degradation D of each component that can describe the entire steady-state process of the unit. j The mean is obtained by calculating the mean as shown in the following formula.

[0068]

[0069] Where T represents the steady-state operating time of the unit from start-up to shutdown.

[0070] (2-4): Finally, the health deterioration of the n components of the unit is spatially fused to obtain the fused deterioration degree D of the hydropower unit, thus forming a time series of the performance deterioration trend of the hydropower unit.

[0071]

[0072] Step (3): Construct a hydropower unit degradation trend prediction model based on Healthformer. The model includes three modules: attention, Bi-GRU, and Quantile Loss. First, position encoding and multi-head attention feature extraction are performed on the historical operating parameters of the unit. Then, the extracted attention vector and the historical degradation degree of the unit are residually connected to form the input of the Bi-GRU module. The feedforward layer is used as a regressor and the conditional quantiles are calculated. Finally, the quantile loss of the model is evaluated and the model parameters are updated and optimized accordingly.

[0073] Step (3) specifically includes the following steps:

[0074] (3-1): The historical operating parameters of the unit are encoded with location and multi-head attention features are extracted. Then, the extracted attention vector and the historical degradation degree of the unit are residually connected. This module includes four parts: location encoding, masked multi-head self-attention, global average pooling layer, and feedforward layer.

[0075] The specific steps of location encoding and multi-head attention are as follows:

[0076] (3-1-1): Regarding the operating parameter x wc Encode according to temporal position to obtain temporal position information p = (p1, ..., p2). m ) T And connect p and x in a residual manner. wc .

[0077] (3-1-2): The multi-head attention mechanism uses the Query-Key-Value calculation method and employs scaled dot product attention. Its calculation formula is given by the following formula:

[0078]

[0079] In the formula: Q represents the query matrix, K represents the matrix showing the relevance of the queried information to other information, and V represents the matrix of the queried information. Meanwhile, Where N and M represent the length of the query matrix and the length of the matrix containing the queried information, respectively, and D k and D v Let K (or Q) and V represent the dimensions of matrices K and V, respectively. The calculation formula for the multi-head attention method is as follows:

[0080] MultiHead(Q,K,V)=Concat(head1,...,head H W O

[0081] in, and W O These are all parameter matrices that need to be trained.

[0082] The residuals are fed into a masked multi-head self-attention network for querying and matching. Similarity is calculated using the Query-Key-Value method to form an attention matrix, which is then flattened using a one-dimensional global average pooling layer to form an attention vector.

[0083] x h =MultiHead(x wc +x p )

[0084] (3-1-3): Finally, the flattened attention vector is input into the feedforward fully connected layer FF1, and the vectors are concatenated and merged to form the hidden feature map of the unit's operating conditions. Provide effective feature parameters for the final prediction.

[0085]

[0086] (3-2): Hide the feature map of the operating condition and historical degradation of the unit x D Residual connections are performed; secondly, the time-series characteristics of unit degradation depth x are obtained through forward GRU and backward GRU calculations. gru The calculation formula is as follows:

[0087]

[0088] (3-3): The feedforward layer FF2 is used as a regressor to generate conditional quantile prediction results, and x gru Inputting these values ​​allows for the calculation of probability predictions of unit degradation at different quantiles γ.

[0089] y D，γ =FF2(x gru )

[0090] The quantile loss function is used to calculate the error between the model output and the true value. The formula for calculating the quantile loss is as follows:

[0091]

[0092] Among them, y i and These represent the true value and the predicted value, respectively. The quantile γ typically ranges from [0, 1]. After the model training is complete, the probability predictions when the quantile γ is 0.95 and 0.05 are used as the upper and lower bounds of the prediction model output interval, respectively.

[0093] Example

[0094] To more clearly illustrate the present invention and highlight its advantages, time-series database data from a hydropower station condition monitoring and fault diagnosis system is used to further explain the present invention.

[0095] Step 1: The dataset spans from May 31, 2017 to November 20, 2020. Model inputs include 13-dimensional feature variables such as active power and reactive power, while model outputs include 15 dependent variables such as vibration of each frame and sway of various parts of the main shaft. Data from May 31, 2017 to May 31, 2018 (one year) was used to train and validate the health model. Data from June 1, 2018 to November 20, 2020 was used to generate a time series showing the performance degradation trend of the hydropower units. The experimental data was divided into training and validation sets in an 8:2 ratio, resulting in 29,091 training data points and 7,273 validation data points. Each data point contains 13-dimensional feature inputs and 15-dimensional outputs.

[0096] Since multiple environmental operating parameters affect the vibration and sway of the hydropower unit, a recursive feature elimination method is used to remove useless environmental feature parameters and obtain key operating parameters that are strongly correlated with the unit's vibration and sway, thereby improving the accuracy of the unit's single-component health model. The relevant operating parameters of six key unit components—upper guide bearing, lower guide bearing, water guide bearing, upper frame vibration, lower frame, and top cover—are calculated and solved under different sensing channels.

[0097] The recursive feature elimination algorithm uses random forest regression as the learner; it selects 5-fold cross-validation as the cross-validation model; and the evaluation metric is the coefficient of determination R. 2 The minimum number of features to select and the number of features to recursively delete are set to 4 and 1, respectively. The random forest model consists of 20 decision trees, sampled with replacement, and the mean squared error (MSE) is selected as the evaluation metric for the decision trees. The random seed is 42.

[0098] Table 1 presents the experimental results of ranking the importance features on a dataset from a hydropower station. It can be seen that all oscillation parameters are strongly correlated with the cooling water inlet temperature, which in turn is related to seasonality, indicating that the oscillation values ​​of the hydropower unit exhibit seasonal variations. The unit's oscillation parameters are strongly correlated with common operating parameters studied, such as head, active power, reactive power, and guide vane opening. The unit's oscillation is not strongly correlated with operating conditions such as excitation current and excitation voltage.

[0099] Table 1. Ranking of the importance of operating condition parameters (smaller values ​​indicate higher importance).

[0100]

[0101] Step 2: Using the feature inputs and outputs obtained in the RFECV stage as the data foundation, a random forest fitting model is built. First, the operating condition parameter data is used as input to run the trained health status index model based on random forest, obtaining 15 single-channel regression results. Based on the Euclidean distance of each component's sensing channel, the performance degradation trends of six key components—upper and lower, water guide bearings, upper and lower frames, and the top cover—are obtained. Time-mean fusion processing is performed on each component based on the steady-state operation process. Finally, spatial fusion is performed on the degradation degree of each component to obtain... Figure 2 The time series showing the combined performance degradation trend of the hydropower units is presented. Figure 2 It can be seen that the overall deterioration degree of the unit gradually increased from 0 to 0.4 over two and a half years, showing an upward trend over time. This indicates that the deterioration degree of the unit gradually worsened during operation. The slight decrease in the overall deterioration degree of the unit after maintenance reflects the accuracy of the health status index model proposed in this invention.

[0102] Step 3: Five variables—temperature, active power, reactive power, guide vane opening, and head—are used as input parameters for the unit operating conditions in the Healthformer model. The Healthformer structure is as follows: Figure 3 As shown. Since the temperatures of the upper guide inlet, lower guide inlet, and water guide inlet show a consistent trend (i.e., consistent with the ambient temperature), the upper guide inlet temperature was chosen as the representative temperature variable. For these operating parameters, due to the lack of data with identical time intervals, the nearest interpolation principle was used to interpolate the data, forming the final experimental data for the Healthformer model. The first 80% of the resulting dataset was used for model training, and the last 20% was used for model testing. The Healthformer model's multi-head attention head and dimension were set to 8, the number of hidden neurons in the bidirectional GRU network was 8, the number of neurons in the two feedforward layers was set to 8 and 32 respectively, and the final output layer dimension was 9, containing 9 different quantiles: [0.005, 0.025, 0.05, 0.1, 0.5, 0.9, 0.95, 0.975, 0.995]. Figure 4 It displays the actual value of the unit's deterioration degree, the prediction model range, the midpoint of the prediction range, and the relative error between the midpoint of the prediction range and the actual value. Figure 4 As can be seen, the proposed Healthformer model maintains a sufficiently small interval width while exhibiting a small difference between the predicted interval center value and the actual unit degradation value, with a relatively uniform distribution. The QQ plot, by calculating the PIT value, visually demonstrates whether the cumulative distribution function of the predicted values ​​follows a uniform distribution, thus validating the model's reliability. Figure 5 This is a QQ chart for predicting the degradation trend range provided in an embodiment of the present invention. For example... Figure 5As shown, the PIT values ​​of the Healthformer model are uniformly distributed around the diagonal (i.e., the theoretical distribution), and almost all PIT points are located near the 5% confidence boundary. This indicates that the prediction results of the proposed model follow a uniform distribution and have strong reliability.

[0103] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for predicting the performance degradation trend of hydropower units considering operating parameters, characterized in that, The method includes: S1. Obtain historical operating parameters of the hydropower unit and calculate the time series of degradation trends that can reflect the overall health status and deterioration performance of the hydropower unit; S2. Input the historical operating parameters and historical deterioration trend time series of the hydropower unit into the trained probability interval prediction model to obtain the reliability prediction interval of the future deterioration of the hydropower unit. The probability interval prediction model consists of a multi-head attention module, a Bi-GRU module, and a quantile loss module. The multi-head attention module extracts hidden feature maps of hydropower unit operating conditions from historical operating parameters of the hydropower unit. After concatenating the hidden feature maps with the historical degradation residuals of the hydropower unit, the data is input to the Bi-GRU module. The Bi-GRU module obtains the time-series features of hydropower unit degradation depth through forward GRU and backward GRU calculations, and outputs them to the quantile loss module. The quantile loss module inputs the time-series features of hydropower unit degradation depth into a feedforward fully connected layer to calculate the probability prediction results of hydropower unit degradation degree at different quantiles.

2. The method as described in claim 1, characterized in that, Step S1 includes: S11. Input the operating parameters of the hydropower unit during operation into the trained single-channel health state model to obtain the oscillation health state value of each channel of the hydropower unit under the operating conditions at each time. Compare it with the real-time oscillation state parameters of the hydropower unit to obtain the degree of degradation under each channel. The single-channel health state model includes random forest and RFECV. The random forest is used to fit the relationship between the operating parameters and the oscillation state parameters. The RFECV is used to remove non-important operating parameters during the training of random forest. S12. The health degradation of individual components during each steady-state operation of the hydropower unit is fused over time to obtain the health degradation of each component that can describe the entire steady-state process of the hydropower unit. S13. Spatially fuse the health degradation of each component of the hydropower unit to obtain the fused degradation of the hydropower unit, thereby forming a time series of hydropower unit performance degradation trend.

3. The method as described in claim 2, characterized in that, In step S11, the degradation degree DC of each channel is calculated. i，t The calculation formula is as follows: Among them, DC i，t ReLU represents the degradation degree of the hydropower unit at time t under channel i, and ReLU represents the linear rectification function. i，t V represents the actual oscillation state parameters of channel i at time t of the hydropower unit. i，t This represents the vibration and oscillation health status value of channel i of the hydropower unit under operating conditions at time t.

4. The method as described in claim 2, characterized in that, In step S12, the formulas for calculating the health deterioration of each component that describes the entire steady-state process of the hydropower unit are as follows: Where time t represents the start time of a certain steady-state operation process, and D j (t) represents the degree of health degradation of component j at time t of the hydropower unit, T represents the steady-state operating time of the hydropower unit from start-up to shutdown, and D j，t This indicates Euclidean distance.

5. The method as described in claim 2, characterized in that, In step S13, the formula for calculating the time series of performance degradation trend of the spatially integrated hydropower units is as follows: Where n represents the number of components in the hydropower unit, time t represents the start time of a certain steady-state operation process, and D(t) represents the time series of the hydropower unit's performance degradation trend. j (t) represents the degree of health degradation of component j of the hydropower unit at time t.

6. The method according to any one of claims 1 to 5, characterized in that, The multi-head attention module includes a position encoding unit, a first residual connection unit, a masked multi-head self-attention unit, a global average pooling layer, a feedforward layer, and a second residual connection unit. The location encoding unit is used to encode the historical operating parameters of the hydropower unit. The first residual connection unit is used to input the residual connection position encoding and operating parameters to the masked multi-head self-attention unit; The masked multi-head self-attention unit is used to extract multi-head attention features and input the extracted attention vector to the feedforward layer. The feedforward layer is used to connect and merge in series to form a hidden feature map of the hydropower unit's operating conditions; The second residual connection unit is used to connect the hidden feature map and the historical deterioration trend time series of the hydropower unit, and then input them into the Bi-GRU module.

7. A system for predicting the performance degradation trend of hydropower units considering operating parameters, characterized in that, include: Processor and memory; The memory is used to store computer programs or instructions; The processor is configured to execute the computer program or instructions in the memory, such that the method of any one of claims 1 to 6 is performed.

Images