Pumped storage unit state trend monitoring method based on integrated prediction model

Abstract

This invention discloses a method for monitoring the state trend of pumped storage units based on an integrated prediction model, comprising the following steps: 1) improving the self-attention mechanism to extract features from the unit's vibration signals; 2) constructing an integrated prediction model, using the first-layer prediction model to achieve individual prediction of time series data; 3) constructing a weighted linear error correction method by integrating and improving the artemisinin optimization algorithm through multi-strategy integration to fuse the prediction results of the first-layer model; 4) optimizing the hyperparameters of the second-layer prediction model based on the improved artemisinin optimization algorithm to achieve efficient and accurate prediction of the fused results; 5) evaluating the lifespan of the pumped storage unit based on the unit state prediction results. The method proposed in this invention can adaptively adjust the weights of predicted and error values ​​and optimize the hyperparameters of the prediction model, improving the efficiency and accuracy of pumped storage unit state trend prediction, accurately warning of abnormal unit operating conditions, ensuring safe and stable unit operation, and improving the overall efficiency of the power station.

Description

Technical Field

[0001] This application belongs to the field of pumped storage unit status trend detection technology, and more specifically, relates to a method for monitoring the status trend of pumped storage units based on an integrated prediction model. Background Technology

[0002] Pumped storage power stations, due to their unique "peak shaving and valley filling" characteristics, play a vital role in regulating load, promoting energy conservation in the power system, and maintaining the safety and stability of the power grid, and have gradually become an indispensable and effective means of power system control. To ensure the safe and stable operation of pumped storage units and improve their overall performance, conducting research on monitoring technologies for the status trends of pumped storage units is crucial for enhancing their intelligent maintenance capabilities.

[0003] Currently, pumped storage units widely employ data-driven prediction methods such as regression analysis, support vector machines, neural networks, and attention mechanisms. However, single deep learning models based on data-driven approaches often blindly discard parts of the prediction discrepancy, leading to a lack of information about the prediction target. To ensure the safe and stable operation of pumped storage units, their vibration characteristics are typically studied, and health monitoring is conducted based on these characteristics. However, due to the sensitivity limitations of vibration sensors, subtle changes in the unit's structure are generally not captured, resulting in insufficient detection accuracy. Summary of the Invention

[0004] In view of the above-mentioned defects or improvement needs of the existing technology, this application provides a method for monitoring the status trend of pumped storage units based on an integrated predictive model. The purpose is to solve the technical problem that the detection accuracy of the existing pumped storage unit monitoring technology based on vibration data is insufficient.

[0005] To achieve the above objectives, this application provides a method for monitoring the state trend of pumped storage units based on an integrated predictive model, which specifically includes the following steps:

[0006] Step 1: Based on the improvement of the self-attention mechanism, a residual connection structure is introduced to improve the stability of the self-attention mechanism, and the improved self-attention mechanism is used to complete the feature extraction of the unit vibration signal to obtain the state feature sequence.

[0007] Step 2: Construct an integrated prediction model. Based on three basic sub-models—bidirectional long short-term memory network, bidirectional gated recurrent unit network, and bidirectional temporal convolutional network—build the first layer of the integrated prediction model. Use the first layer prediction model to achieve individual prediction of time series.

[0008] Step 3: Improve the artemisinin optimization algorithm through multi-strategy integration, construct a weighted linear error correction method, and complete the fusion of the prediction results of the first-layer sub-model in Step 2;

[0009] Step 4: Based on the improved artemisinin optimization algorithm in Step 3, optimize the hyperparameters of the second-layer prediction model. Using the fusion result from Step 3 as input, use the optimized second-layer prediction model to achieve efficient and accurate prediction of the fusion result from Step 3. The second-layer prediction model is a quantile regression random forest model.

[0010] Step 5: Evaluate the lifespan of the pumped storage unit based on the unit status prediction results in Step 4.

[0011] Furthermore, step 1 improves the self-attention mechanism as follows:

[0012] (1-1) Replace the randomly initialized weight matrix with three 1s. 1. Convolution and the introduction of nonlinear features enhance the expressiveness of the improved self-attention mechanism;

[0013] (1-2) A residual connection structure is introduced to improve the stability of the improved self-attention mechanism. The feature sequence extracted by the improved self-attention mechanism is as follows:

[0014]

[0015] In the formula, B represents the output feature of the improved self-attention mechanism, and s x f represents the input feature sequence. a Let Q and K represent the weight parameters to be learned, and D represent the feature matrices. K This represents the channel signals Q and K, where K... T Indicates the transpose of K;

[0016] (1-3) The improved self-attention mechanism is used to extract features from the unit vibration signal and obtain its state feature sequence C.

[0017] Further, in step 2, the first layer of the integrated prediction model is constructed, and the three basic sub-models of the first layer prediction model, namely the bidirectional long short-term memory network, the bidirectional gated recurrent unit network, and the bidirectional temporal convolutional network, are trained using the input state feature sequence to obtain three sets of output results.

[0018] Furthermore, step 3 includes the following sub-steps:

[0019] (3-1) Improve the artemisinin optimization algorithm through multi-strategy ensemble, specifically as follows:

[0020] A differential mutation strategy is adopted, which uses mutation to enable individuals to escape local optima and achieve position update. The position update process is as follows:

[0021]

[0022] In the formula, H represents the position of the i-th mutated individual in the n-th iteration, and H represents the differential scaling factor. This represents the position of the Q-th individual in the nth iteration. This represents a randomly generated distinct integer in the range [1, Q] that is not equal to Q, where Q represents the total size;

[0023] A dynamic crossover mechanism is adopted, which simulates the dynamic characteristics of the solution space and adjusts the search process using the current iteration number and the maximum iteration number. The new solution component Y is:

[0024]

[0025] In the formula, and These represent two solution components, The positive tuning parameter that controls the decay rate is denoted by N, where N represents the total number of iterations and n represents the nth iteration.

[0026] (3-2) Construct a weighted linear error correction method, divide the prediction results into two parts: point prediction and residual prediction, use the improved artemisinin optimization algorithm to find the optimal weights of the two parts of prediction values, and merge them to form a new state sequence;

[0027] In the weight allocation process, the MAPE function is used as the fitness function to measure the difference between the predicted value and the true value. The MAPE expression is as follows:

[0028]

[0029] In the formula, pred represents point prediction, and residual represents residual prediction. The value represents the true value of the sample, h represents the number of samples, and e1 and e2 represent the weights of the point prediction and the residual prediction, respectively.

[0030] By continuously iterating and updating the improved artemisinin optimization algorithm population position, the optimal weights of point prediction and residual prediction are calculated using the fitness function, and a new state sequence is formed by fusing the optimal weights.

[0031] Furthermore, the fused state sequence X in (3-2) is as follows:

[0032]

[0033] In the formula, Let represent the predicted value of the j-th sequence in the t-th sub-model. Let represent the weight of the j-th predicted value in the t-th sub-model, and M represent the total number of samples.

[0034] Furthermore, step 4 applies the improved artemisinin optimization algorithm to the parameter optimization task of the quantile regression random forest model. Specifically, the optimization targets are the maximum depth of the model decision tree, the number of leaf nodes in the decision tree, and the number of variables involved in splitting in the leaf nodes. The optimization process continuously iterates and updates the population position based on the improved artemisinin optimization algorithm, realizes the joint search of hyperparameters, and coordinates the prediction interval and learning ability of the quantile regression random forest, ultimately achieving a more efficient and accurate prediction of the fusion state sequence.

[0035] Furthermore, step 5 calculates the mean absolute error, root mean square error, and coefficient of determination based on the predicted value of the final output of the model, thereby realizing the life assessment of the pumped storage unit.

[0036] Overall, the technical methods conceived in this application have the following advantages compared with the prior art:

[0037] (1) This invention adopts a differential mutation strategy for the first time to solve the problem that the artemisinin optimization algorithm is prone to getting trapped in local optima; at the same time, a dynamic crossover mechanism is introduced to enable the artemisinin optimization algorithm to adaptively balance exploration and development, thereby achieving faster convergence and higher accuracy.

[0038] (2) This invention is the first to consider constructing a weighted linear error correction method, using an improved artemisinin optimization algorithm to calculate the optimal weights of the point prediction value and the residual prediction value, thereby achieving adaptive fusion of prediction results, fully preserving the prediction information of each sub-model in the first layer of the integrated model, and improving the predictability of the unit status trend.

[0039] (3) This invention is the first to construct an integrated prediction model, which integrates a bidirectional long short-term memory network, a bidirectional gated recurrent unit network, a bidirectional temporal convolutional network and a quantile regression random forest model. Using the constructed weighted linear error correction method, the weights of the three output results of the first layer of the integrated model are adaptively assigned to form a new state sequence. With the new state sequence as input, the optimized quantile regression random forest model is used to complete the efficient prediction of the new state sequence, thereby obtaining accurate prediction results of the unit state development trend. This significantly improves the prediction accuracy of the unit state trend, overcomes the theoretical problem of the difficulty in effectively realizing unit state early warning, and provides reliable theoretical guidance and technical support for the implementation of intelligent state maintenance of pumped storage units. Attached Figure Description

[0040] Figure 1 A flowchart of the pumped storage unit status trend monitoring method provided by the present invention;

[0041] Figure 2 This is the state sequence change curve of the pumped storage unit in this embodiment of the invention.

[0042] Figure 3 This is a diagram of the integrated prediction model architecture in an embodiment of the present invention;

[0043] Figure 4 Flowchart of the improved artemisinin optimization algorithm provided by this invention;

[0044] Figure 5 This is a curve showing the predicted state development trend of the pumped storage unit in an embodiment of the present invention. Detailed Implementation

[0045] The present invention will now be described in detail with reference to the accompanying drawings.

[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. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0047] This invention uses the status trend monitoring of a pumped storage unit in a domestic power station as an analysis case. Figure 1 The flowchart shown is a method for monitoring the status trend of pumped storage units provided by the present invention, which specifically includes the following steps:

[0048] Step 1: Based on the improvement of the self-attention mechanism, complete the feature extraction of the unit vibration signal.

[0049] (1-1) Replace the randomly initialized weight matrix with three 1s. 1. Convolution and the introduction of nonlinear features enhance the expressiveness of the improved self-attention mechanism;

[0050] (1-2) A residual connection structure is introduced to improve the stability of the improved self-attention mechanism. The feature sequence extracted by the improved self-attention mechanism is as follows:

[0051]

[0052] In the formula, B represents the output feature of the improved self-attention mechanism, and s x f represents the input feature sequence. a Let Q and K represent the weight parameters to be learned, and D represent the feature matrices. K This represents the channel signals Q and K, where K... T This represents the transpose of K.

[0053] (1-3) Based on the improved self-attention mechanism, feature extraction is performed on the unit vibration signal to obtain its output feature C. A case study is taken as the early warning system status of a pumped storage unit in a domestic power station. Figure 2 The figure shows the sequence of changes in the unit's initial state.

[0054] Step 2: Construct an integrated forecasting model and use the first-layer sub-model to achieve individual forecasting of time series.

[0055] (2-1) Based on three basic sub-models (bidirectional long short-term memory network, bidirectional gated recurrent unit network, and bidirectional temporal convolutional network), the first layer of the integrated prediction model is constructed.

[0056] (2-2) Input the extracted state feature sequence C into the first layer of the ensemble model to obtain three sets of output results.

[0057] Step 3: Improve the artemisinin optimization algorithm through multi-strategy integration, construct a weighted linear error correction method, and complete the fusion of the prediction results of the first-layer sub-model.

[0058] (3-1) A differential mutation strategy is adopted to prevent getting trapped in local optima by generating mutated individuals.

[0059] The differential mutation strategy utilizes mutations to allow an individual to escape local optima, thereby achieving position updates. The position update process is as follows:

[0060]

[0061] In the formula, H represents the position of the i-th mutated individual in the n-th iteration, and H represents the differential scaling factor. This represents the position of the Q-th individual in the nth iteration. This represents a different integer randomly generated within the range [1, Q] that is not equal to Q (where Q represents the total size).

[0062] (3-2) A dynamic crossover mechanism is adopted, which simulates the dynamic characteristics of the solution space and adjusts the search process using the current iteration number and the maximum iteration number. The new solution component Y is:

[0063]

[0064] In the formula, and These represent two solution components, This represents the positive tuning parameter that controls the decay rate.

[0065] (3-3) Construct a weighted linear error correction method, divide the prediction results into two parts: point prediction and residual prediction, use the improved artemisinin optimization algorithm to find the optimal weights of the two predicted values, and merge them to form a new state sequence.

[0066] In the weight allocation process, the MAPE function was used as the fitness function to measure the difference between the predicted and actual values. The MAPE expression is as follows:

[0067]

[0068] In the formula, pred represents point prediction, and residual represents residual prediction. The value represents the true value of the sample, h represents the number of samples, and e1 and e2 represent the weights of the point prediction and the residual prediction, respectively.

[0069] By iteratively updating and improving the artemisinin optimization algorithm to determine the population position, the optimal weights for point prediction and residual prediction are calculated using a fitness function. Based on these weights, a new state sequence is formed by fusing the results. The fused state sequence X is:

[0070]

[0071] In the formula, Let represent the predicted value of the j-th sequence in the t-th sub-model. Let represent the weight of the j-th predicted value in the t-th sub-model, and M represent the total number of samples.

[0072] like Figure 3 The diagram shows the architecture of the integrated prediction model. It can be seen that the system integrates a bidirectional long short-term memory network, a bidirectional gated recurrent unit network, a bidirectional temporal convolutional network, and a quantile regression random forest model. The feature sequence is used as the input of the integrated model, and the output is the prediction result of the development trend of the state sequence.

[0073] like Figure 4 The flowchart shown is the improved algorithm for artemisinin optimization. It can be seen that it integrates differential mutation strategy and dynamic crossover mechanism, which overcomes the problem of the population getting stuck in local optima during the search process and realizes the algorithm converges to the global optimum.

[0074] Step 4: Based on the improved artemisinin optimization algorithm, optimize the hyperparameters of the second-layer prediction model to achieve efficient and accurate prediction of the fusion results.

[0075] (4-1) The improved artemisinin optimization algorithm was applied to the parameter optimization task of the quantile regression random forest model. Specifically, the optimization targets were the maximum depth of the decision tree, the number of leaf nodes in the decision tree, and the number of variables involved in the splitting of leaf nodes. This optimization process was based on the improved artemisinin optimization algorithm, which continuously iteratively updated the population position during the optimization process to achieve joint search of hyperparameters. This mechanism coordinated the adjustment of the prediction interval and learning ability of the quantile regression random forest, ultimately achieving more efficient and accurate prediction of the fusion state sequence.

[0076] (4-2) The optimized model is used to fit the fusion results to obtain the final prediction results.

[0077] like Figure 5 The figure shows the predicted state development trend curve of the pumped storage unit. As can be seen from the results, the final predicted curve fits the original state sequence change curve well, fully verifying the effectiveness of the above-mentioned state trend prediction method.

[0078] Step 5: Evaluate the lifespan of the pumped storage unit based on the unit status prediction results.

[0079] Two evaluation metrics were used to assess the unit. The first is the root mean square error (RMSE), which reflects the deviation between the predicted and actual values, i.e.:

[0080]

[0081] In the formula, L is the number of samples; l represents the l-th sample; y l This represents the actual value of the l-th sample; This represents the final predicted value for the l-th sample.

[0082] The second type is the coefficient of determination R. score 2 Its value ranges from a maximum of 1 to a minimum of 0. A value closer to 1 indicates better model performance; a value closer to 0 indicates worse model performance. Specifically:

[0083]

[0084] In the formula, This represents the average value of the actual sample.

[0085] The above content is readily understood by those skilled in the art. The above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for monitoring the state trend of pumped storage units based on an integrated predictive model, characterized in that, Specifically, the steps include the following: Step 1: Based on the improvement of the self-attention mechanism, a residual connection structure is introduced to improve the stability of the self-attention mechanism, and the improved self-attention mechanism is used to complete the feature extraction of the unit vibration signal to obtain the state feature sequence. Step 2: Construct an integrated prediction model. Based on three basic sub-models—bidirectional long short-term memory network, bidirectional gated recurrent unit network, and bidirectional temporal convolutional network—build the first layer of the integrated prediction model. Use the first layer prediction model to achieve individual prediction of time series. Step 3: Improve the artemisinin optimization algorithm through multi-strategy integration, construct a weighted linear error correction method, and complete the fusion of the prediction results of the first-layer sub-model in Step 2; Step 4: Based on the improved artemisinin optimization algorithm in Step 3, optimize the hyperparameters of the second-layer prediction model. Using the fusion result from Step 3 as input, use the optimized second-layer prediction model to achieve efficient and accurate prediction of the fusion result from Step 3. The second-layer prediction model is a quantile regression random forest model. Step 5: Evaluate the lifespan of the pumped storage unit based on the unit status prediction results in Step 4.

2. The method for monitoring the state trend of pumped storage units based on an integrated predictive model according to claim 1, characterized in that, Step 1 improves the self-attention mechanism as follows: (1-1) Replace the randomly initialized weight matrix with three 1s.

1. Convolution and the introduction of nonlinear features enhance the expressiveness of the improved self-attention mechanism; (1-2) A residual connection structure is introduced to improve the stability of the improved self-attention mechanism. The feature sequence extracted by the improved self-attention mechanism is as follows: ； In the formula, B represents the output feature of the improved self-attention mechanism, and s x f represents the input feature sequence. a Let Q and K represent the weight parameters to be learned, and D represent the feature matrices. K This represents the channel signals Q and K, where K... T Indicates the transpose of K; (1-3) The improved self-attention mechanism is used to extract features from the unit vibration signal and obtain its state feature sequence C.

3. The method for monitoring the state trend of pumped storage units based on an integrated predictive model according to claim 1, characterized in that, Step 2 involves constructing the first layer of the integrated prediction model. The three basic sub-models of the first layer prediction model—a bidirectional long short-term memory network, a bidirectional gated recurrent unit network, and a bidirectional temporal convolutional network—are trained using the input state feature sequence to obtain three sets of output results.

4. The method for monitoring the state trend of pumped storage units based on an integrated predictive model according to claim 1, characterized in that, Step 3 includes the following sub-steps: (3-1) Improve the artemisinin optimization algorithm through multi-strategy ensemble, specifically as follows: A differential mutation strategy is adopted, which uses mutation to enable individuals to escape local optima and achieve position update. The position update process is as follows: ； In the formula, H represents the position of the i-th mutated individual in the n-th iteration, and H represents the differential scaling factor. This represents the position of the Q-th individual in the nth iteration. This represents a randomly generated distinct integer in the range [1, Q] that is not equal to Q, where Q represents the total size; A dynamic crossover mechanism is adopted, which simulates the dynamic characteristics of the solution space and adjusts the search process using the current iteration number and the maximum iteration number. The new solution component Y is: ； In the formula, and These represent two solution components, The positive tuning parameter that controls the decay rate is denoted by N, where N represents the total number of iterations and n represents the nth iteration. (3-2) Construct a weighted linear error correction method, divide the prediction results into two parts: point prediction and residual prediction, use the improved artemisinin optimization algorithm to find the optimal weights of the two parts of prediction values, and merge them to form a new state sequence; In the weight allocation process, the MAPE function is used as the fitness function to measure the difference between the predicted value and the true value. The MAPE expression is as follows: ； In the formula, pred represents point prediction, and residual represents residual prediction. The value represents the true value of the sample, h represents the number of samples, and e1 and e2 represent the weights of the point prediction and the residual prediction, respectively. By continuously iterating and updating the improved artemisinin optimization algorithm population position, the optimal weights of point prediction and residual prediction are calculated using the fitness function, and a new state sequence is formed by fusing the optimal weights.

5. The method for monitoring the state trend of pumped storage units based on an integrated predictive model according to claim 4, characterized in that, The fused state sequence X in (3-2) is as follows: ； In the formula, Let represent the predicted value of the j-th sequence in the t-th sub-model. Let represent the weight of the j-th predicted value in the t-th sub-model, and M represent the total number of samples.

6. The method for monitoring the state trend of pumped storage units based on an integrated predictive model according to claim 1, characterized in that, Step 4 applies the improved artemisinin optimization algorithm to the parameter optimization task of the quantile regression random forest model. Specifically, the optimization targets are the maximum depth of the model decision tree, the number of leaf nodes in the decision tree, and the number of variables involved in splitting in the leaf nodes. The optimization process is based on the improved artemisinin optimization algorithm, which continuously iterates and updates the population position during the optimization process to achieve joint search of hyperparameters. This collaboratively adjusts the prediction interval and learning ability of the quantile regression random forest, ultimately achieving a more efficient and accurate prediction of the fusion state sequence.

7. The method for monitoring the state trend of pumped storage units based on an integrated predictive model according to claim 1, characterized in that, Step 5 calculates the mean absolute error, root mean square error, and coefficient of determination based on the predicted value of the final output of the model, thereby realizing the life assessment of the pumped storage unit.

Images