Method for predicting temperature of key positions of steam turbine based on multi-model fusion and PSO optimization

Abstract

This invention belongs to the field of power equipment condition monitoring and intelligent operation and maintenance technology. It discloses a multi-model fusion method for predicting generator critical location temperatures based on PSO optimization. The key steps are: constructing a dynamic adaptive fusion model of multiple prediction models based on exponentially weighted moving average; using particle swarm optimization (PSO) to optimize the fusion parameters of the fusion model constructed in step S4, obtaining the optimal fusion parameters as the parameter configuration of the final fusion model; and using the final fusion model to predict the temperature of the generator critical location over the next m steps, outputting the prediction results. This invention's multi-model fusion method for predicting generator critical location temperatures based on PSO optimization builds a multi-model fusion framework based on the exponentially weighted moving average method and uses the particle swarm optimization (PSO) algorithm to optimize the hyperparameters in the fusion calculation process, obtaining the best prediction effect for generator critical location temperatures over multiple time steps under the optimal parameters.

Description

Technical Field

[0001] This invention belongs to the field of power generation equipment condition monitoring and intelligent operation and maintenance technology, specifically involving a method for predicting the temperature of key locations of generators based on PSO optimization and multi-model fusion. Background Technology

[0002] Large steam turbine generators are core equipment in thermal power plants and nuclear power plants, and their operating status directly affects the reliability, stability, and safety of the power system. Due to prolonged operation in high-temperature, high-pressure, and high-speed rotating environments, steam turbine generators are susceptible to factors such as temperature, leading to insulation aging, stator failures, and rotor failures. Generator temperature is a crucial monitoring indicator during operation; therefore, predicting the temperature at critical locations can prevent problems such as delayed warnings and untimely fault diagnosis. This facilitates intelligent operation and maintenance and predictive maintenance, thereby avoiding significant economic losses due to unexpected downtime.

[0003] Research on trend prediction of generator critical location temperatures falls under the category of time series prediction problems, a common type of regression problem. Currently, existing condition monitoring methods mainly fall into two categories: traditional statistical model-based prediction methods and emerging artificial intelligence prediction methods based on machine learning or deep learning models. In recent years, researchers have increasingly used machine learning-related artificial intelligence methods, such as Long Short-Term Memory Networks (LSTM), to predict generator critical location temperatures, achieving significantly higher prediction accuracy than traditional methods. Research on the fusion of multiple prediction models is relatively limited. Some researchers have used a fusion of LSTM and XGBoost algorithms to predict generator stator winding temperatures, achieving prediction results superior to either LSTM or XGBoost alone. Other researchers have proposed a multi-step prediction method for the stator bar outlet water temperature difference by fusing LSTM and ARIMA models; this ARMA-LSTM model algorithm outperforms either ARIMA or LSTM alone.

[0004] Current research on multi-prediction model fusion algorithms has two main limitations: (1) The proposed fusion algorithm is limited to the fusion between two model algorithms and is rarely extended to prediction algorithms that fuse multiple models.

[0005] (2) The fusion parameters have a great influence on the final effect of the fusion model. Previous studies have only superficially analyzed the influence of hyperparameters and have not used intelligent optimization methods to optimize the fusion parameters in order to obtain prediction results with smaller errors and higher accuracy. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a multi-model fusion method for predicting the temperature at critical locations of generators based on PSO (Particle Swarm Optimization). This method takes the temperature at critical locations of large steam turbine generators as the research object, builds a multi-model fusion framework based on the Exponentially Weighted Moving-Average (EWMA) method, and uses the PSO algorithm to optimize the hyperparameters in the fusion calculation process to obtain the best prediction effect of the generator's critical location temperature at multiple time steps under the optimal parameters.

[0007] The technical solution to achieve the above objective is: a multi-model fusion method for predicting the temperature at critical locations of a generator based on PSO optimization, comprising the following steps: S1, acquire historical temperature monitoring data at key locations of the generator to form a time series dataset; S2, Select n different single prediction models, and apply the historical temperature monitoring data collected in step S1 to predict the future for each prediction model. m Multi-step time series prediction with a step size; S3, Calculate the global root mean square error of each prediction model over m prediction steps. RMSE m ,by RMSE m For predictive evaluation indicators, RMSE m The specific formula is as follows: In the formula: for The true value of the sample at time step. The predicted value output by the model. T The number of samples; Will RMSE m The smallest predictive model is labeled 0, and the remaining predictive models are labeled as follows: ; to facilitate the use of calculation formulas in subsequent construction of fusion models; S4, Construct a dynamic adaptive fusion model of multiple prediction models based on exponentially weighted moving average; S5. The particle swarm optimization algorithm is used to optimize the fusion parameters of the fusion model constructed in step S4, and the optimal fusion parameters are used as the parameter configuration of the final fusion model. S6. Using the final fusion model obtained in step S5, predict the temperature of the key locations of the generator in the next m steps and output the prediction results.

[0008] The above-mentioned multi-model fusion method for predicting generator critical location temperature based on PSO optimization includes the following steps in step S4: All n Global root mean square error in each prediction model RMSE m The smallest predictive model is labeled as 0, and its fusion weight is set to 1; the remaining predictive models are labeled as... When fusing the prediction results at t time steps, the EWMA framework is introduced to construct the fusion weights for the remaining prediction models: In the formula: ; For the first The fusion weight change rate parameter of each EWMA model, and ; For the first The fusion weights of the models are t The observations at time 1 are used to dynamically increase or decrease the fusion weights at that time. ; The calculation formula is as follows: In the formula, u It is a step function. ; ; For the first i The model to be merged is in the first... t Time before k The fusion loss function for the step is: Time model corresponding The step-weighted squared error is calculated as follows: In the formula: For the first The model at time step The predicted value; For the corresponding true value; k For the number of prediction steps, and 1 ≤k≤m ; They are respectively the front In each prediction step, step 1 to step 2 The fusion weights of the steps; the fusion loss function corresponding to the model with the best prediction performance is denoted as... The fusion loss functions for other models are denoted as follows: , ..., ; Based on the fusion weights of each prediction model, a dynamic adaptive fusion model of multiple prediction models based on exponentially weighted moving average is constructed, and its calculation formula is as follows: In the formula: for t Predicted values ​​after time-mapping; The model with the best prediction performance for a single overall system is... RMSE m The smallest predicted value; For the remaining first i The predicted values ​​of a prediction model.

[0009] The above-mentioned multi-model fusion method for predicting generator critical location temperature based on PSO optimization includes the following steps in step S5: S5.1, Determine the set of fusion parameters to be optimized, including: (previous parameters) k In each prediction step, step 1 to step 2 k Weights of the previous k steps and prediction weights The value range is [0,1]; and there are n-1 fusion weight change rates. , , ..., The value range is [0,1]; S5.2, Define based on the fusion parameters to be optimized. For particles in 3D space, determine the initial values ​​of algorithm parameters such as particle population size, position, velocity, and inertia weight, and initialize the particle swarm; S5.3, Define the prediction steps for different fusion loss functions. The root mean square error of the prediction performance is evaluated under different parameters to be optimized. As a fitness function; S5.4, for each particle in the particle swarm, at a specific prediction step... k Under the fusion parameters to be optimized, based on the fusion loss function formula (Formula (5)) and the fusion calculation formulas in step S4 (Formula (3)-Formula (7)), dynamic adaptive fusion calculation is performed on the individual prediction values ​​of multiple prediction models to obtain the fusion function value of each particle under the set of fusion parameters to be optimized. That is, the future after integration m Multi-step time series prediction values ​​under a step size; S5.5, For each particle in the particle swarm, based on the predicted value obtained in step S5.5, calculate the fitness function value of each particle under the set of fusion parameters to be optimized. ; S5.6 According to the particle swarm optimization algorithm calculation formula, for each particle, the fusion parameters to be optimized in its corresponding fusion model are updated using a linear decreasing inertia weight strategy. S5.7, in the prediction steps k Then, continuously iterate and repeat steps S5.4 to S5.6 until convergence, thereby obtaining the optimal fusion parameters when the prediction effect is optimal. as well as and its corresponding m Step Predicted Value Save its corresponding prediction error ; S5.8, Traverse the prediction steps k In 1 to m Repeat steps S5.4 to S5.7 for each value between these intervals to obtain the desired result. k Find the prediction error and corresponding optimal fusion parameters that achieve the best prediction performance under the given values; select the fusion loss function and prediction steps that minimize the prediction error. And the corresponding optimal fusion parameters.

[0010] In the aforementioned multi-model fusion method for predicting generator critical location temperature based on PSO optimization, step S5.6 involves the particle swarm optimization algorithm being used as the velocity update formula. and position update formula : In the formula: i =1, 2, ..., ; d =1, 2, ..., D ; h This represents the current iteration number; The inertial weight reflects the ability to inherit previous velocities; P id The first optimal position for an individual d Dimensional components; P gd The optimal position for the group d Dimensional components; c 1 and c 2 is the weighting factor; r 1 and r 2 is a random number; The particle velocity and position are updated using a linearly decreasing inertia weight strategy, as shown in the following formula: In the formula: This is the maximum inertial weight value, which is also the initial inertial weight value; This is the minimum inertia weight value, which is also the inertia weight ending value when the maximum number of iterations is reached; This represents the current iteration number; max This represents the maximum number of iterations; the iteration search is terminated when the iteration converges.

[0011] The above-mentioned multi-model fusion method for predicting generator critical location temperature based on PSO optimization, wherein the particle swarm optimization algorithm has a particle population size of 50-200, a maximum number of iterations of 100-500, and a weight factor. c 1 and c The value of 2 is in the range of [1, 2].

[0012] The above-mentioned multi-model fusion method for predicting generator critical location temperature based on PSO optimization includes a single prediction model comprising a long short-term memory network model, an extreme learning machine model, a recurrent neural network model, a convolutional neural network model, and / or an autoregressive integral moving average model.

[0013] The above-mentioned method for predicting the temperature of critical locations of a generator based on PSO optimization and multi-model fusion includes the stator core, stator winding, rotor winding and / or bearings.

[0014] The above-mentioned method for predicting generator critical location temperature using multi-model fusion based on PSO optimization has a prediction step number m ranging from 1 to 12.

[0015] The generator critical location temperature prediction method based on PSO optimization multi-model fusion of the present invention has the following technical effects: (1) Using the EWMA model framework, a model was constructed that extends to n The multi-prediction model multi-step adaptive fusion formula of the model can realize the fusion of multiple prediction models and the prediction effect is better than that of two models. (2) The PSO algorithm is used to optimize the fusion parameters, and the influence of the EWMA model weights and the fusion formula weights is analyzed in depth. At the same time, the following measures are taken: m Predict step size, select 1 to 2. m Between k The process involves iterating through multiple hyperparameters. Optimizing these hyperparameters can further improve prediction performance. Attached Figure Description

[0016] Figure 1 The flowchart shows the generator critical location temperature prediction method based on PSO optimization multi-model fusion according to the present invention. Figure 2 This is a schematic diagram of the PSO optimization iteration process in the example; Figure 3The graph shows the optimization performance of the optimal fusion parameter model compared to the ELM model for different prediction steps k. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0018] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0019] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0020] Please see Figure 1 The present invention provides a multi-model fusion method for predicting generator critical location temperature based on PSO optimization, comprising the following steps: S1, acquire historical temperature monitoring data at key locations of the generator to form a time series dataset; S2, Select n different single prediction models, and apply the historical temperature monitoring data collected in step S1 to predict the future for each prediction model. m Multi-step time series prediction with a step size; S3, calculate the global root mean square error of each prediction model over m prediction steps; S4, Construct a dynamic adaptive fusion model of multiple prediction models based on exponentially weighted moving average; S5. The particle swarm optimization algorithm is used to optimize the fusion parameters of the fusion model constructed in step S4, and the optimal fusion parameters are used as the parameter configuration of the final fusion model. S6. Using the final fusion model obtained in step S5, predict the temperature of the key locations of the generator in the next m steps and output the prediction results.

[0021] The key technologies of this invention, namely, the multi-model fusion method for predicting generator critical location temperature based on PSO optimization, are a dynamic adaptive fusion model of multiple prediction models based on EWMA (step S4) and a time-series prediction method based on the dynamic adaptive fusion of multiple models using the particle swarm optimization (PSO algorithm). These are described in detail below.

[0022] I. Constructing a dynamic adaptive fusion model of multiple prediction models based on EWMA: 1.1 Construction of the fusion model: The Exponentially Weighted Moving Average (EWMA) model is a further optimization of the Simple Moving Average (SMA) and Weighted Moving Average (WMA) models. This model uses an exponentially decreasing weighting method, assigning higher weights to data closer to the current time point and exponentially decreasing weights to data further away from the current time point. Through this mechanism, the EWMA model can more accurately reflect changes in recent data while gradually reducing the influence of distant historical data on current results, thus achieving more flexible responsiveness and smoothing effects in time series forecasting.

[0023] The calculation formula for the EWMA model is: (1) In formula (1): t For time; for t The fitted value at time; These are the weighting coefficients for the observed values; for t The observed value at time; for The fitted value at time t.

[0024] This invention uses multiple EWMA model calculation formulas to construct a dynamic adaptive fusion model of multiple prediction models. n Model fusion to predict the future m Taking time-series data of a step size as an example, the fusion model proposed in this invention will be introduced.

[0025] First, define the current time step. t accumulated by sensors p The time series data at each time step is used as the input to the fusion model, denoted as . The model aims to predict the future. Step sensor data, i.e. , for The actual value of a moment. (The time step) The corresponding sample is denoted as .

[0026] Suppose we now have nThis invention uses the root mean square error (RMSE) as an error evaluation metric in a prediction model. RMSE ) Select the optimal model, using m Step RMSE m The function is as follows: (2) In formula (2): for The true value of the sample at time step. The predicted value output by the model. T This represents the number of samples.

[0027] All n Global prediction performance in each prediction model RMSE m The smallest prediction model is labeled 0, and its fusion weight is set to 1. When fusing the prediction results at t time steps, this invention introduces the EWMA framework to construct the remaining prediction models (labeled sequentially as follows). ) fusion weights: (3) In formula (3): ; For the first The fusion weight change rate parameter of each EWMA model, and ; For the first The fusion weights of the models are t The observations at time 1 are used to dynamically increase or decrease the fusion weights at that time. ; The calculation formula is as follows: (4) In equation (4), u It is a step function; For the first i The model to be merged is in the first... t Time before k The fusion loss function for the step is: Time model corresponding The step-weighted squared error is calculated as follows: (5) In formula (5): For the first The model at time step The predicted value; For the corresponding true value; k For the number of prediction steps, and 1 ≤k≤m ; They are respectively the front In each prediction step, step 1 to step 2 The fusion weights of the steps; the fusion loss function corresponding to the model with the best prediction performance is denoted as... The fusion loss functions for other models are denoted as follows: , ..., ; u It can be represented as: (6) Based on the fusion weights of each prediction model (the fusion weights constructed based on formula (3)), a dynamic adaptive fusion model of multiple prediction models based on exponential weighted moving average is constructed, and its calculation formula is as follows: (7) In equation (7): for t Predicted values ​​after time-mapping; for t The model with the best overall prediction performance at a single moment is... RMSE m The smallest predicted value; for t The remaining time i The predicted values ​​of the prediction model; for t Time of the first i Dynamic fusion weights of a prediction model; 1.2 Implementation of the integration model characteristics and adaptive functionality: Based on the complete set of formulas constructed from the above fusion model, the adaptive fusion model of multiple prediction models has the following characteristics: 1) When ( When all ) are 1, it indicates the global prediction effect. RMSE m The smallest model is in the first place. Step-by-step local prediction effect WSE The optimal EWMA fit value is... It is still close to 1. According to formula (7), the final prediction result is... The weighting is relatively large. This is because the generator temperature data is at different time steps. and The data trend characteristics are similar, and the global prediction effect is good. RMSE m Minimum model at time step It may still have the best predictive performance, so the fusion model is reasonable.

[0028] 2) When ( When both are 0, the corresponding Approaching 0, meaning the model with the best global prediction performance is in the th... The local prediction performance is worse than that of other prediction models. In this case, the denominator of formula (7) can be approximated as: n -1, the fused function value can be approximated as: (8) That is, the fused function value is the average of the prediction values ​​of all the best non-global prediction models at the current time.

[0029] 3) When )middle j The value is 0. middle j a value (1≤ j ≤ n -1) is close to 0, that is The top-performing model for single-time overall prediction k Step prediction performance is inferior to others. j For each prediction model, the denominator of formula (7) can be approximated as: j The fused function value can be approximated as: (9) According to formula (9), the fused function value is the current time. j The average of the predicted values ​​of the non-single overall best prediction model. Observing the above formula, it can be seen that case 2) is actually case 3). Special circumstances as follows.

[0030] 4) As can be seen from formula (7), the weights of different prediction models are dynamically changing. The changing pattern is as follows: when the prediction effect of the single overall best prediction model is better than the other prediction models at a certain moment, the weights of other prediction models that are worse than the single overall best prediction model are reduced; at some moments, the single overall best prediction model may be worse than other models, and the weights of other models at these moments are increased, thereby realizing the dynamic adaptive adjustment of the weights of each model.

[0031] II. The time-series prediction method using PSO-optimized multi-model dynamic adaptive fusion also employs the particle swarm optimization algorithm to find the optimal fusion parameters of the constructed fusion model, selecting the fusion loss function and prediction steps corresponding to the minimum prediction error. and the corresponding optimal fusion parameters and the future input of this set of parameters. m Step Predicted Value A multi-step time series data prediction method based on PSO optimization and dynamic adaptive fusion of multiple models was developed.

[0032] 2.1 Selection of fusion parameters to be optimized: As can be seen from the fusion model construction process, the fusion parameters are influenced by the prediction weights. and EMWA model hyperparameters The impact is also affected by the number of prediction weight steps. k Impact. Therefore, two types of fusion parameters can be selected: 1) Before Prediction weights of steps ,common The parameters, since the weight normalization in the fusion calculation process eliminates the constraint that the sum is 1, are thus... The values ​​can all take values ​​from zero to infinity, and the range of values ​​is the same. For the convenience of optimization calculation, the range of values ​​is set to [0,1]. 2) Rate of change of fusion weights , , ..., ,common n- One parameter, defined by the EWMA model formula, has a value range of [0,1]. For the number of prediction steps k Considering the cumulative error effect of the prediction steps, in order to ensure reliable prediction accuracy, the number of prediction steps is... m The possible values ​​are generally limited, corresponding to the ones that can be obtained. k Since the number of values ​​is also finite, we can directly use the traversal method to find them. k (1≤ k ≤m) Optimization 2.2 Using PSO to optimize the fusion model: Particle Swarm Optimization (PSO) is an intelligent optimization algorithm that simulates the social behavior of flocks of birds or schools of fish. Its basic idea is to find the optimal solution in a complex solution space through information sharing and cooperation among individuals within the group.

[0033] The core of the particle swarm optimization algorithm is the velocity update formula. and position update formula : (10) In formula (10): i =1, 2, ..., ; d =1, 2, ..., D ; h This represents the current iteration number; The inertial weight reflects the ability to inherit previous velocities; P id The first optimal position for an individual d Dimensional components; P gd The optimal position for the group d Dimensional components; c 1 and c2 is the weighting factor; r 1 and r 2 is a random number.

[0034] This invention employs the PSO algorithm to optimize the parameters of the fusion model. It mainly includes the following aspects: 1) Construction of particle swarm motion space: Considering the parameters to be optimized, k + n -1, constructed according to the PSO algorithm definition. k + n -1 dimension space.

[0035] 2) Fitness function selection: The global prediction performance of the fusion model is used as the fitness evaluation index, and the root mean square error (RMSE) is selected as the fitness function. The formula is the same as that in formula (2).

[0036] 3) Inertia weight Iterative strategy To effectively balance the global exploration and local exploitation capabilities of the particle swarm optimization (PSO) algorithm, a linear descending weight (LDW) strategy is adopted. This strategy causes the descending weight to decrease linearly with the number of iterations, thus enabling the algorithm to possess strong global search capabilities in the early stages of iteration to explore new regions, and strong local search capabilities in the later stages of iteration to finely search the region near the optimal solution. The formula is as follows: (11) In equation (11): This is the maximum inertial weight value, which is also the initial inertial weight value; This is the minimum inertia weight value, which is also the inertia weight ending value when the maximum number of iterations is reached; This represents the current iteration number; max This represents the maximum number of iterations; the iteration search is terminated when the iteration converges.

[0037] By adopting the technical solution proposed in this invention, it is helpful to accurately extract key spatiotemporal features from motor operation data, significantly improve the ability to predict the temperature at key locations of the generator, thereby enabling early warning of the operating trend of key motor parameters, effectively reducing the risk of motor system downtime caused by faults, reducing maintenance costs, and improving the reliability and safety of equipment operation.

[0038] III. Please refer to Figure 2 and Figure 3To evaluate the model's performance, this invention conducts a case study based on actual operating data from a large generator in a motor factory. This data is collected by temperature sensors installed on the generator set, recording various temperature values, including those of the stator core. The sampling frequency is once per hour, and the data duration covers one year.

[0039] Taking the stator core temperature of a steam turbine generator as the research object, this study used stator temperature monitoring data collected by the equipment from May 2023 to May 2024 as the dataset (the monitoring system sampled at 1-hour intervals, totaling 8760 data points) for experimental verification. The data from the first 11 months of the dataset were used as the test set (totaling 8040 data points), and the data from the last month (totaling 720 data points) were used as the test set for model training.

[0040] Two LSTM models with different initialization states and one ELM model were selected for dynamic adaptive fusion and particle swarm optimization between pairwise and three models to obtain stator core prediction values ​​for the next 5 hours. The prediction performance of stator temperature was evaluated using three prediction error metrics: root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE). m Step prediction error index RMSE m The expression is shown in formula (2), MAE m and MAPE m The expression is as follows: (12) (13) In equations (12) and (13): For the true value, The predicted values ​​obtained from the fusion model. It is a first-order norm. T The number of samples; m To predict the number of steps, here m =5.

[0041] 3.1 Constructing a dynamic adaptive fusion model for multiple prediction models based on EWMA: For the fusion of three models, when n=3, the fusion calculation formula (7) simplifies to: (14) In equation (14), for t Predicted values ​​after time-mapping; x t for tHistorical temperature data at any given time; For ELM model in t Individual predicted value at any given time; For the first LSTM model in t Individual predicted value at any given time; For the second LSTM model in t Individual predicted value at any given time; For the first model to be fused (LSTM1) in t The dynamic fusion weight at each time step, with a value range of [0,1]; For the second model to be fused (LSTM2) t The dynamic fusion weight at any given time has a value range of [0,1].

[0042] For pairwise model fusion, the ELM model is selected to be fused with LSTM model 1 and LSTM model 2 respectively. The fusion calculation formula (7) is simplified to: (15) In equation (15), for t Predicted values ​​after time-mapping; for t ELM model prediction at time 10:00 for t LSTM model prediction at time; For the model to be fused (LSTM) in t time Dynamic fusion weights.

[0043] The physical meaning of formula (15): When hour, (Using ELM prediction entirely); when hour, (Using LSTM prediction exclusively); when When the two are used, the weighted average is calculated.

[0044] 3.2 A time series forecasting method using PSO-optimized multi-model dynamic adaptive fusion: Set c1=1.49445, c2=1.49445, ω_max=0.9, ω_min=0.4, V_max=0.5, V_min=0.5. Set the population size to 100 and the maximum number of generations to 200. Use PSO to... k The fusion parameters were optimized when the fusion parameters were 1, 2, 3, 4, and 5.

[0045] 3.3 Application Effects: Table 1 shows the prediction errors of the three prediction models, ELM, LSTM1 and LSTM2, over a 5-hour period.

[0046] Table 1. Prediction error of each individual prediction model: As shown in Table 1, ELM has the best prediction performance. Therefore, ELM and LSTM1 and ELM and LSTM2 are fused together to perform pairwise fusion.

[0047] When fusing three models, ELM-LSTM1-LSTM2, taking a prediction step count of k=3 as an example, there are five parameters to be optimized: α1, α2, ω1, ω2, and ω3. The PSO optimization iteration process is as follows: Figure 2 As shown.

[0048] At this point, different prediction steps k Below, the optimization performance of its optimal fusion hyperparameter model compared to the ELM model is shown in [reference needed]. Figure 3 As shown: Depend on Figure 3 It can be seen that, k When the value is 3, the overall prediction effect is optimal.

[0049] Using two-model fusion methods from existing literature and multiple model fusion methods proposed in this invention, PSO optimization was performed using ELM, LSTM1, and LSTM2 for pairwise fusion and three-model fusion, respectively. The prediction performance relative to the best of the three models, ELM, is shown in Table 2.

[0050] Table 2 shows the optimization effect of prediction error under different prediction methods: As shown in Table 2, the method of this invention significantly improves prediction performance compared to existing two-model fusion methods in the literature. For two-model and three-model fusion methods based on this invention, the three prediction errors (RMSE, MAE, and MAPE) are all significantly better than those of the single optimal ELM model, and also significantly better than the three prediction errors of methods based on the literature. Furthermore, it can be observed that when applying the method of this invention with appropriate multiple prediction models, increasing the number of prediction models can achieve even better prediction results.

[0051] This invention enables more accurate multi-time-step prediction of temperatures at critical locations in steam turbine generators. Multi-time-step prediction of temperatures at critical locations in steam turbine generators by fusing multiple prediction models presents two technical challenges: 1) Previous research primarily focused on fusing two prediction models; how to use the EWMA model framework to derive a construction formula that can be extended to the fusion of multiple prediction models? 2) Previous fusion model construction processes have involved limited analysis of fusion calculation parameters; how to select appropriate fusion parameters and optimize them using optimization techniques to achieve the best prediction results? This invention solves both of these technical challenges.

[0052] In summary, the generator critical location temperature prediction method based on PSO optimization of the present invention builds a multi-model fusion framework based on the exponential weighted moving average method and uses the particle swarm optimization algorithm to optimize the hyperparameters in the fusion calculation process, so as to obtain the best prediction effect of generator critical location temperature at multiple time steps under the optimal parameters.

[0053] Those skilled in the art should recognize that the above embodiments are merely illustrative of the present invention and are not intended to limit the present invention. Any variations or modifications to the above embodiments that are within the spirit and essence of the present invention will fall within the scope of the claims of the present invention.

Claims

1. A method for predicting the temperature at critical locations of a generator based on PSO optimization and multi-model fusion, characterized in that, Includes the following steps: S1, acquire historical temperature monitoring data at key locations of the generator to form a time series dataset; S2, Select n different single prediction models, and apply the historical temperature monitoring data collected in step S1 to predict the future for each prediction model. m Multi-step time series prediction with a step size; S3, Calculate the global root mean square error of each prediction model over m prediction steps. RMSE m ,by RMSE m For predictive evaluation indicators, RMSE m The specific formula is as follows: In the formula: for The true value of the sample at time step. The predicted value output by the model. T The number of samples; Will RMSE m The smallest predictive model is labeled 0, and the remaining predictive models are labeled as follows: ; S4, Construct a dynamic adaptive fusion model of multiple prediction models based on exponentially weighted moving average; S5. The particle swarm optimization algorithm is used to optimize the fusion parameters of the fusion model constructed in step S4, and the optimal fusion parameters are used as the parameter configuration of the final fusion model. S6. Using the final fusion model obtained in step S5, predict the temperature of the key locations of the generator in the next m steps and output the prediction results.

2. The method for predicting generator critical location temperature based on PSO optimization and multi-model fusion according to claim 1, characterized in that, Step S4 specifically includes the following process: All n Global root mean square error in each prediction model RMSE m The smallest predictive model is labeled as 0, and its fusion weight is set to 1; Label the remaining prediction models as When fusing the prediction results at t time steps, the EWMA framework is introduced to construct the fusion weights for the remaining prediction models: In the formula: ; For the first The fusion weight change rate parameter of each EWMA model, and ; For the first The fusion weights of the models are t The observations at time 1 are used to dynamically increase or decrease the fusion weights at that time. ; The calculation formula is as follows: In the formula, u It is a step function. ; ; For the first i The model to be merged is in the first... t Time before k The fusion loss function for the step is: Time model corresponding The step-weighted squared error is calculated as follows: In the formula: For the first The model at time step The predicted value; For the corresponding true value; k For the number of prediction steps, and 1 ≤k≤m ; They are respectively the front In each prediction step, step 1 to step 2 The fusion weights of the steps; the fusion loss function corresponding to the model with the best prediction performance is denoted as... The fusion loss functions for other models are denoted as follows: , ..., ; Based on the fusion weights of each prediction model, a dynamic adaptive fusion model of multiple prediction models based on exponentially weighted moving average is constructed, and its calculation formula is as follows: In the formula: for t Predicted values ​​after time-mapping; The model with the best prediction performance for a single overall system is... RMSE m The smallest predicted value; For the remaining first i The predicted values ​​of a prediction model.

3. The method for predicting generator critical location temperature based on PSO optimization and multi-model fusion according to claim 2, characterized in that, Step S5 specifically includes the following process: S5.1, Determine the set of fusion parameters to be optimized, including: (previous parameters) k In each prediction step, step 1 to step 2 k Weights of the previous k steps and prediction weights The value range is [0,1]; and there are n-1 fusion weight change rates. , , ..., The value range is [0,1]; S5.2, Define based on the fusion parameters to be optimized. For particles in 3D space, determine the initial values ​​of algorithm parameters such as particle population size, position, velocity, and inertia weight, and initialize the particle swarm; S5.3, Define the prediction steps for different fusion loss functions. The root mean square error of the prediction performance is evaluated under different parameters to be optimized. As a fitness function; S5.4, for each particle in the particle swarm, at a specific prediction step... k Under the fusion parameters to be optimized, based on the fusion loss function formula and the fusion calculation formulas in step S4, dynamic adaptive fusion calculation is performed on the individual prediction values ​​of multiple prediction models to obtain the fused function value of each particle under the set of fusion parameters to be optimized. That is, the future after integration m Multi-step time series prediction values ​​under a step size; S5.5, For each particle in the particle swarm, based on the predicted value obtained in step S5.5, calculate the fitness function value of each particle under the set of fusion parameters to be optimized. ; S5.6 According to the particle swarm optimization algorithm calculation formula, for each particle, the fusion parameters to be optimized in its corresponding fusion model are updated using a linear decreasing inertia weight strategy. S5.7, in the prediction steps k Then, continuously iterate and repeat steps S5.4 to S5.6 until convergence, thereby obtaining the optimal fusion parameters when the prediction effect is optimal. as well as and its corresponding m Step Predicted Value Save its corresponding prediction error ; S5.8, Traverse the prediction steps k In 1 to m Repeat steps S5.4 to S5.7 for each value between these intervals to obtain the desired result. k Find the prediction error and corresponding optimal fusion parameters that achieve the best prediction performance under the given values; select the fusion loss function and prediction steps that minimize the prediction error. And the corresponding optimal fusion parameters.

4. The method for predicting generator critical location temperature based on PSO optimization and multi-model fusion according to claim 3, characterized in that, In step S5.6, the particle swarm algorithm uses the velocity update formula. and position update formula : In the formula: i =1, 2, ..., ; d =1, 2, ..., D ; h This represents the current iteration number; The inertial weight reflects the ability to inherit previous velocities; P id The first optimal position for an individual d Dimensional components; P gd The optimal position for the group d Dimensional components; c 1 and c 2 is the weighting factor; r 1 and r 2 is a random number; The particle velocity and position are updated using a linearly decreasing inertia weight strategy, as shown in the following formula: In the formula: This is the maximum inertial weight value, which is also the initial inertial weight value; This is the minimum inertia weight value, which is also the inertia weight ending value when the maximum number of iterations is reached; This represents the current iteration number; max This represents the maximum number of iterations; the iteration search is terminated when the iteration converges.

5. The method for predicting generator critical location temperature based on PSO optimization using multi-model fusion as described in claim 4, characterized in that, In the particle swarm optimization algorithm, the particle population size is set to 50-200, the maximum number of iterations is set to 100-500, and the weight factor is... c 1 and c The value of 2 is in the range of [1, 2].

6. The method for predicting generator critical location temperature based on PSO optimization and multi-model fusion according to claim 1, characterized in that, The single prediction model includes a long short-term memory network model, an extreme learning machine model, a recurrent neural network model, a convolutional neural network model, and / or an autoregressive integral moving average model.

7. The method for predicting generator critical location temperature based on PSO optimization and multi-model fusion according to claim 1, characterized in that, The key components of the generator include the stator core, stator windings, rotor windings, and / or bearings.

8. The method for predicting generator critical location temperature based on PSO optimization and multi-model fusion according to claim 1, characterized in that, The prediction step m ranges from 1 to 12.

Images