A maintenance-free moisture absorber state prediction method and system based on multi-model fusion

Abstract

This invention provides a method and system for predicting the condition of a maintenance-free desiccant based on multi-model fusion, relating to the field of power equipment condition prediction technology. The method includes the following steps: data acquisition, data preprocessing, feature engineering, parallel prediction using multiple models, adaptive fusion, and condition assessment. This invention integrates three models: LSTM, Random Forest, and XGBoost, fully leveraging the advantages of each model: the LSTM model excels at capturing time-series features and can accurately capture the long-term trend of desiccant saturation; the Random Forest model excels at handling nonlinear relationships and can effectively uncover the influence of environmental features on the desiccant's condition; and the XGBoost model has strong generalization ability and can accurately handle the mapping relationship between operating condition features and fault risks. Through the synergistic effect of multiple models, the prediction accuracy is improved, the prediction error is small, and the high-precision prediction requirements of maintenance-free desiccant are met.

Description

Technical Field

[0001] This invention relates to the field of power equipment condition prediction technology, and in particular to a method and system for predicting the condition of a maintenance-free dehumidifier based on multi-model fusion. Background Technology

[0002] Transformer oil conservator dehumidifiers are crucial auxiliary devices for ensuring the safe and stable operation of transformers. Their core function is to absorb moisture from the air inside the transformer oil conservator, preventing moisture from entering the transformer and causing insulation degradation, equipment aging, or even malfunctions. Currently, the condition monitoring of maintenance-free dehumidifiers in power systems mainly relies on two technical approaches: one is the traditional manual inspection method, where maintenance personnel visually observe the color change of the internal silica gel to determine the desiccant's moisture absorption status; the other is a prediction method based on a single machine learning model, which collects limited operational data and uses a single model to predict the condition. Meanwhile, while some existing intelligent dehumidifiers possess basic monitoring functions, they often suffer from poor data acquisition accuracy, insufficient regeneration stability, and a lack of a comprehensive prediction mechanism, making it difficult to achieve truly maintenance-free management.

[0003] Traditional manual inspections are time-consuming, subjective, and unable to monitor desiccant status in real time, resulting in blind spots and an inability to predict desiccant saturation trends, making it difficult to develop maintenance plans in advance. Prediction methods based on single machine learning models struggle to simultaneously capture the temporal, environmental, and operational characteristics of the dehumidifier, exhibiting poor adaptability to complex and ever-changing power operating environments, limited prediction accuracy, and a lack of dynamic optimization capabilities. They cannot adjust prediction strategies in real time based on model performance and lack interpretability of prediction results, hindering fault diagnosis and model optimization. These shortcomings all contribute to untimely dehumidifier maintenance, thereby threatening the safe and stable operation of transformers. Therefore, this invention proposes a maintenance-free dehumidifier status prediction method and system based on multi-model fusion to address the problems existing in the prior art. Summary of the Invention

[0004] To address the aforementioned issues, this invention proposes a multi-model fusion-based method and system for predicting the state of a maintenance-free desiccant. It integrates three models: LSTM, Random Forest, and XGBoost, leveraging the strengths of each: the LSTM model excels at capturing temporal features and accurately captures the long-term trend of desiccant saturation; the Random Forest model excels at handling nonlinear relationships and effectively mines the influence of environmental features on the desiccant's state; and the XGBoost model has strong generalization capabilities and can accurately handle the mapping relationship between operating state characteristics and fault risks. Through the synergistic effect of multiple models, prediction accuracy is improved, prediction errors are minimized, and the high-precision prediction requirements for maintenance-free desiccant are met.

[0005] To achieve the objectives of this invention, the invention is implemented through the following technical solution: a maintenance-free dehumidifier state prediction method based on multi-model fusion, comprising the following steps:

[0006] S1: Data Acquisition: Real-time acquisition of dehumidifier operation data through multi-source sensors, including humidity data, temperature data, pressure data, and vibration data;

[0007] S2: Data preprocessing: The collected raw data is cleaned, outlier handling is performed, missing value imputation is performed, and data normalization is performed to obtain a standardized dataset;

[0008] S3: Feature Engineering: Extracting temporal, statistical, and environmental features from standardized datasets to construct multi-dimensional feature vectors;

[0009] S4: Multi-model parallel prediction: The feature vectors are input into the LSTM model, the random forest model and the XGBoost model respectively for parallel prediction, and the prediction results of the three sub-models are obtained.

[0010] S5: Adaptive Fusion: Based on the prediction performance of each sub-model on historical data, dynamically calculate the weight coefficients, weight and fuse the prediction results of the three sub-models, and output the final prediction result.

[0011] S6: Condition Assessment: Based on the fusion prediction results, assess the current condition of the desiccant, predict the desiccant saturation and remaining service life.

[0012] A further improvement is made in S2, where outlier handling is performed during data preprocessing. The criteria identify and replace outliers, and the range of outlier determination is: , The mean of the data. The standard deviation of the data is given; missing values ​​are filled using linear interpolation, and when more than 3 consecutive sampling points are missing, the mean of the previous and next 10 sampling points is used for filling; data normalization uses the Min-Max method, mapping each feature data to the [0,1] interval, and the normalization formula is:

[0013] ,

[0014] in, The normalized data has a value range of [0,1] and is used to eliminate the influence of different dimensions of data to make the feature data comparable; x is the original data to be normalized, that is, a single feature data after data collection, cleaning, outlier processing, and missing value imputation. It is the minimum value in the dataset for this feature, that is, the minimum value among all the original data for the same feature; It is the maximum value in the dataset for this feature, that is, the maximum value among all the original data for the same feature.

[0015] Further improvements are made in S3, where the time-series features include the sliding window mean, sliding window variance, trend slope, and periodic components. The mean, variance, and standard deviation are calculated using a 24-hour sliding window, the trend slope is calculated using the least squares method, and the periodic components are extracted using Fourier transform. The statistical features include the maximum value, minimum value, median, quartiles, skewness, and kurtosis, and the above statistical features of each data point over the past 7 days are calculated. The environmental features include ambient temperature, ambient humidity, atmospheric pressure, and seasonal factors, with the seasonal factors corresponding to 1 / 2 / 3 / 4 for spring / summer / autumn / winter, respectively.

[0016] Further improvements are made in S4, where the LSTM model is used to process temporal features. Its model structure includes an input layer, two LSTM hidden layers, a fully connected layer, and an output layer. The input time window is set to 24-72 hours, the number of neurons in the hidden layer is 64-128, the dropout rate is controlled at 0.2-0.3, the model training uses the Adam optimizer with a learning rate of 0.001, the number of training epochs is 100, an early stopping mechanism is used to prevent overfitting, and the output is the predicted value of desiccant saturation for the next 24-72 hours.

[0017] Further improvements are made in S4, where a random forest model is used to process environmental features. The number of decision trees is set to 100-200, the maximum depth is 10-20 layers, the minimum number of split samples is 5-10, the feature sampling ratio is 0.6-0.8, and the Gini coefficient is used as the splitting criterion. The model mainly explores the nonlinear relationship between environmental features and desiccant saturation and outputs the predicted value of desiccant saturation.

[0018] Further improvements are made in S4, where the XGBoost model is used to process the running state features. Its learning rate is 0.01-0.1, the number of trees is 100-300, the maximum depth is 3-8 layers, the L2 regularization coefficient is 0.1-1.0, and the mean squared error (MSE) is used as the loss function. It mainly captures the mapping relationship between running state features and fault risk, and outputs fault risk-related predicted values.

[0019] A further improvement is made in S5, where the adaptive fusion weight coefficients are calculated as follows: based on the root mean square error (RMSE) of each sub-model over the most recent N time windows, the weights are dynamically adjusted using an exponentially weighted moving average method, where N is a positive integer. The weight calculation formula is as follows:

[0020] ,

[0021] The fusion calculation formula is as follows:

[0022] ;

[0023] in, is the weight coefficient of the i-th sub-model, where i takes the values ​​1, 2, and 3, corresponding to the LSTM model, random forest model, and XGBoost model, respectively; exp is the natural exponential function; This is an adjustment coefficient used to adjust the degree of influence of prediction error on the weighting coefficient, with a value of 2.0. is the root mean square error of the i-th sub-model over the most recent N time windows, used to measure the prediction accuracy of the i-th sub-model. The smaller the error, the larger the weighting coefficient. The summation symbol is 'j', which takes values ​​of 1, 2, and 3, corresponding to the three sub-models. It is used to sum the exponential terms of the three sub-models. This is the final prediction result after multi-model fusion; These are the weight coefficients of the LSTM model; These are the independent prediction results of the LSTM model; These are the weight coefficients of the random forest model; These are the independent prediction results from the random forest model; Y represents the weight coefficients of the XGBoost model; Y_XGB represents the independent prediction results of the XGBoost model.

[0024] Further improvements are made in the following aspects: In S6, the state assessment includes: desiccant saturation assessment, which divides saturation into four levels: 0-30% is normal, 30%≤saturation<60% is slightly saturated, 60%≤saturation<80% is moderately saturated, and saturation≥80% is heavily saturated; remaining service life prediction, which predicts the number of days required for the desiccant to reach the 80% saturation threshold; and failure risk classification, where remaining service life>14 days is low risk, 7 days<remaining service life≤14 days is medium risk, and remaining service life≤7 days is high risk.

[0025] Further improvements include: an online model update step: incremental training of the three sub-models and fusion strategy is performed every 7 days using the latest collected operational data, and the model parameters are updated; the prediction error of each sub-model is monitored in real time, and model retraining is triggered when the average prediction error of a sub-model exceeds a preset threshold for 30 consecutive days; the model training adopts a transfer learning strategy, first pre-training with a publicly available power equipment operation dataset, and then fine-tuning with field data.

[0026] A maintenance-free dehumidifier condition prediction system based on multi-model fusion includes a data acquisition module, a data preprocessing module, a feature engineering module, a multi-model prediction module, a fusion decision module, and a condition assessment module. The data acquisition module is used to collect humidity, temperature, pressure, and vibration data related to the dehumidifier's operation in real time through multi-source sensors. The data preprocessing module is used to perform data cleaning, outlier handling, missing value imputation, and data normalization on the collected raw data to output a standardized dataset. The feature engineering module is used to extract temporal features, statistical features, and environmental features from the standardized dataset to construct a multi-dimensional feature vector.

[0027] The multi-model prediction module incorporates LSTM, random forest, and XGBoost models to perform parallel predictions on feature vectors and output independent prediction results for each sub-model. The fusion decision module dynamically calculates weight coefficients based on the historical prediction performance of each sub-model and performs weighted fusion of the prediction results to output the final prediction result. The state assessment module assesses the current operating state of the dehumidifier, predicts the desiccant saturation and remaining service life, and classifies the failure risk level based on the final prediction result.

[0028] The beneficial effects of this invention are as follows:

[0029] 1. This invention integrates three models: LSTM, Random Forest, and XGBoost, giving full play to the advantages of each model: the LSTM model is good at capturing time-series features and can accurately capture the long-term trend of desiccant saturation; the Random Forest model is good at handling nonlinear relationships and can effectively explore the influence of environmental features on the state of the dehumidifier; the XGBoost model has strong generalization ability and can accurately handle the mapping relationship between operating state features and fault risks. Through the synergistic effect of multiple models, the prediction accuracy is improved and the prediction error is small, meeting the high-precision prediction requirements of maintenance-free dehumidifiers.

[0030] 2. The feature engineering of this invention extracts multi-dimensional features of time series, statistics, and environment, comprehensively covering various factors affecting the state of the dehumidifier and avoiding prediction bias caused by a single feature. The adaptive fusion mechanism dynamically adjusts the weight coefficients according to the real-time prediction performance of each sub-model through the exponential weighted moving average method to ensure that the prediction results are always optimal. At the same time, through the online update mechanism and retraining trigger conditions, the model can adapt to the complex and ever-changing power operation environment, avoid model performance degradation, and maintain stable prediction results in the long term.

[0031] 3. This invention can not only accurately predict desiccant saturation, but also predict remaining service life and classify fault risk levels, achieving saturation trend prediction 7-14 days in advance. This facilitates maintenance personnel to formulate maintenance plans in advance, truly realizing intelligent management of maintenance-free dehumidifiers. At the same time, the prediction results of the three sub-models can be analyzed independently, which can clearly locate the key factors affecting the state of the dehumidifier, facilitate fault diagnosis and model optimization, improve the practicality and operability of the method, reduce maintenance costs, and ensure the safe and stable operation of transformers. Attached Figure Description

[0032] Figure 1 This is a flowchart of the present invention;

[0033] Figure 2 This is a comparison chart of the ablation experiment results of the present invention;

[0034] Figure 3 This is a comparison diagram of the multi-model fusion architecture of the present invention;

[0035] Figure 4 This is a time series comparison chart of the predictive performance of the present invention. Detailed Implementation

[0036] To enhance understanding of the present invention, the present invention will be further described in detail below with reference to embodiments. These embodiments are only used to explain the present invention and do not constitute a limitation on the scope of protection of the present invention.

[0037] Example 1

[0038] according to Figure 1 , 2 As shown in Figures 3 and 4, this embodiment proposes a method and system for predicting the state of a maintenance-free dehumidifier based on multi-model fusion, including the following steps:

[0039] S1: Data Acquisition. Real-time data acquisition of the dehumidifier's operation is achieved through multi-source sensors deployed on and around the dehumidifier body. The sampling frequency is set to once every 10 minutes to ensure the real-time nature and continuity of the data. Specific data acquired includes: desiccant humidity (measurement range 0-100%RH, accuracy ±2%RH) and ambient humidity data collected by a humidity sensor; desiccant temperature (measurement range -20℃-80℃, accuracy ±0.5℃) and ambient temperature data collected by a temperature sensor; internal pressure of the dehumidifier (measurement range 0-1MPa, accuracy ±0.01MPa) data collected by a pressure sensor; and transformer vibration (measurement range 0-50Hz, accuracy ±0.1Hz) data collected by a vibration sensor. All sensors are industrial-grade to ensure data acquisition accuracy.

[0040] S2: Data Preprocessing. Data preprocessing employs a step-by-step approach to ensure data quality:

[0041] Data cleaning: Remove abnormal data caused by sensor malfunctions and signal interference, with a focus on removing data that is outside the sensor's measurement range and data that is obviously illogical (such as negative humidity of the desiccant).

[0042] Outlier handling: using The criterion for identifying outliers is when the data exceeds a certain threshold. Scope (of which) The mean of the data. When the value is less than the standard deviation of the data, it is considered an outlier and replaced with the mean of the five sampling points before and after the outlier.

[0043] Missing value handling: For missing data, linear interpolation is preferred for filling. If more than 3 consecutive sampling points are missing, the average of the 10 sampling points before and after is used for filling to avoid missing data affecting subsequent processing.

[0044] Data normalization: The Min-Max normalization method is used to map each feature data to the [0,1] interval. The normalization formula is as follows: Where x is the original data, This is the minimum value of the feature data. This is the maximum value of the feature data, ensuring that data from different dimensions are comparable.

[0045] S3: Feature Engineering. Extract multi-dimensional features from the preprocessed, standardized dataset and construct multi-dimensional feature vectors.

[0046] Time series characteristics: The mean, variance, and standard deviation of each data point are calculated using a 24-hour sliding window. The trend slope of each data point is calculated using the least squares method. The periodic components of the data are extracted (periodic features are extracted using Fourier transform) to comprehensively reflect the time variation pattern of the data.

[0047] Statistical characteristics: Calculate the maximum, minimum, median, quartiles (upper quartile Q1, lower quartile Q3), skewness, and kurtosis of each data point over the past 7 days to clearly reflect the distribution characteristics and dispersion of the data;

[0048] Environmental characteristics: Real-time data of ambient temperature, humidity and atmospheric pressure are collected, and seasonal factors are calculated based on the date (spring / summer / autumn / winter correspond to 1 / 2 / 3 / 4 respectively) to fully consider the impact of the external environment on the dehumidifier's operating status;

[0049] Feature selection: The extracted features were selected using the analysis of variance (ANOVA) method. Features with too small variance and weak correlation with the dehumidifier status were removed, while features with strong correlation were retained to construct the final multi-dimensional feature vector.

[0050] S4: Multi-model parallel prediction. The constructed multi-dimensional feature vectors are simultaneously input into three sub-models for parallel prediction. The model configurations are as follows:

[0051] LSTM Model: The model structure is as follows: Input layer (feature dimension × time step) → LSTM layer 1 (128 neurons) → Dropout layer (0.2) → LSTM layer 2 (64 neurons) → Dropout layer (0.2) → Fully connected layer (32 neurons) → Output layer (1 neuron). The input time window is set to 48 hours, and the output is the predicted desiccant saturation value for the next 24 hours. The model is trained using the Adam optimizer with a learning rate of 0.001 and 100 training epochs. Early stopping is used to prevent overfitting.

[0052] Random Forest Model: The number of decision trees is set to 150, the maximum depth is 15 layers, the minimum number of split samples is 5, the feature sampling ratio is 0.7, and the Gini coefficient is used as the splitting criterion. It mainly deals with the nonlinear relationship between environmental features and desiccant saturation and outputs the predicted value of desiccant saturation.

[0053] The XGBoost model has a learning rate of 0.05, 200 trees, a maximum depth of 6 layers, an L2 regularization coefficient of 0.5, and uses mean squared error (MSE) as the loss function. It mainly processes the mapping relationship between operating state characteristics (pressure, vibration) and failure risk, and outputs failure risk-related predicted values.

[0054] S5: Adaptive Fusion. The prediction results of the three sub-models are integrated using a dynamic weight fusion method. The specific process is as follows:

[0055] Error calculation: Calculate the root mean square error (RMSE) of each sub-model over the most recent 30 time windows (each time window is 10 minutes). The smaller the RMSE, the higher the model prediction accuracy.

[0056] Weight Calculation: The weight coefficients of each sub-model are calculated using the exponentially weighted moving average method. The calculation formula is as follows: ,in Let be the weights of the i-th model. Let be the root mean square error of the i-th model. The adjustment coefficient (with a value of 2.0) ensures that the smaller the error, the greater the weight of the model.

[0057] Fusion calculation: The final prediction result is a weighted sum of the prediction results of the three sub-models, calculated using the following formula: ,in The prediction results of the LSTM model. These are the prediction results from the random forest model. The results are predictions from the XGBoost model.

[0058] S6: Status Assessment. Based on the fusion prediction results, comprehensively assess the current operating status of the dehumidifier:

[0059] Desiccant saturation assessment: The desiccant saturation S is divided into four levels: 0-30% is the normal state, at which point the dehumidifier works normally and requires no maintenance; 30%≤S<60% is the slightly saturated state, requiring close monitoring of the dehumidifier's operation; 60%≤S<80% is the moderately saturated state, requiring advance maintenance planning; S≥80% is the heavily saturated state, requiring immediate maintenance.

[0060] Remaining useful life (RUL) prediction: Based on the growth trend of desiccant saturation and combined with historical data, predict the number of days required for the desiccant to reach the 80% saturation threshold, providing a time basis for maintenance decisions;

[0061] Fault Risk Classification: Fault risk levels are classified according to remaining service life (RUL). RUL > 14 days is low risk, 7 days < RUL ≤ 14 days is medium risk, and RUL ≤ 7 days is high risk. Corresponding operation and maintenance suggestions are output for different risk levels.

[0062] Example 2

[0063] according to Figure 1 , 2 As shown in Figures 3 and 4, this embodiment proposes a method and system for predicting the state of a maintenance-free dehumidifier based on multi-model fusion. The difference from Embodiment 1 lies in the model training method and online update mechanism; the remaining steps are the same as in Embodiment 1. The specific improvements are as follows:

[0064] The model training employs a transfer learning strategy to reduce training costs and improve generalization ability: First, pre-training is performed using publicly available power equipment operation datasets (such as the dehumidifier operation dataset in the IEEE PES Power Systems Test Case Archive) to initialize model parameters and give the model basic predictive capabilities. Then, fine-tuning is performed using field-collected dehumidifier operation data to specifically optimize model parameters and adapt the model to the field operating environment. An early stopping mechanism is used during training: training stops when the validation set loss does not decrease for 10 consecutive epochs, effectively preventing overfitting and improving generalization ability.

[0065] Online update mechanism: To ensure that the model can adapt to changes in the field operating environment, the three sub-models and the fusion strategy are incrementally trained every 7 days using the latest collected operating data, and the model parameters are updated so that the model can learn new data features in real time. At the same time, the prediction error of each sub-model is monitored in real time, and the average prediction error of each model is recorded every day. If the average error of a certain sub-model exceeds the preset threshold (RMSE>10%) for 30 consecutive days, the model is retrained and the model parameters are re-optimized to ensure that the model always maintains a high prediction accuracy.

[0066] Example 3

[0067] according to Figure 1 , 2 As shown in Figures 3 and 4, this embodiment proposes a method and system for predicting the state of a maintenance-free dehumidifier based on multi-model fusion. The difference between this embodiment and Embodiment 1 lies in the feature engineering and fusion strategy; the remaining steps are the same as in Embodiment 1. The specific improvements are as follows:

[0068] Feature engineering optimization: Based on Example 1, a frequency domain feature extraction step is added. The collected vibration signal is subjected to Fast Fourier Transform (FFT) to convert the time domain signal into a frequency domain signal, and the main frequency component and energy spectrum features (such as the maximum energy value and energy distribution variance in the frequency domain) are extracted to further mine the dehumidifier operating status information contained in the vibration signal. At the same time, an attention mechanism is introduced to automatically learn the importance weight of each feature, assigning higher weights to important features and lower weights to irrelevant features, thereby improving the effectiveness of the feature vector and further improving the model prediction accuracy.

[0069] Optimization of the fusion strategy: A stacking method is adopted to replace the weighted fusion strategy in Example 1, further improving prediction accuracy. The specific process is as follows: First, LSTM, Random Forest, and XGBoost are used as base models for parallel prediction, obtaining the prediction results of the three base models; then, the prediction results of the three base models are used as input features to construct a new feature vector; finally, logistic regression is used as a meta-model to train and predict the new feature vector, outputting the final fused prediction result. Through secondary fusion, the prediction error is further reduced and the prediction accuracy is improved. Verification shows that this fusion strategy can further reduce the RMSE of saturation prediction by 0.3-0.5 percentage points.

[0070] Validation data:

[0071] The invention underwent a 6-month field verification at a substation. The prediction performance comparison with a single model and traditional methods is shown in the table below:

[0072]

[0073] The results show that the prediction accuracy of the multi-model fusion method of the present invention is significantly better than that of the single model and the traditional method. The prediction error of desiccant saturation is less than 5%, and the prediction error of remaining service life is less than 2 days, which meets the intelligent management requirements of maintenance-free dehumidifiers.

[0074] Meanwhile, to demonstrate the necessity of multi-model fusion, ablation experiments were conducted, and the results are as follows: Figure 2 As shown:

[0075] Performance comparison of different model combinations: The accuracy of a single LSTM model is 82%, the accuracy of a single random forest model is 80%, the accuracy of a single XGBoost model is 79%, while the accuracy of the LSTM+random forest+XGBoost fusion model reaches 91%, which is 9-12 percentage points higher than the single model.

[0076] Comparison of fusion strategies: Adaptive weighted fusion achieved the lowest MAE across all prediction durations, with a 24-hour prediction MAE of 0.10, which is better than average fusion (0.18) and static weighted fusion (0.15).

[0077] Feature input ablation experiment: Removing vibration features resulted in the largest performance decrease (-8%), while removing temporal features resulted in a 6% performance decrease, verifying the necessity of multi-source features.

[0078] In addition, comparative verification was conducted on the multi-model fusion architecture, such as... Figure 3 As shown: Architecture A (early fusion) performs fusion at the feature level and then makes a unified prediction, achieving an accuracy of 82%; Architecture B (mid-term fusion) performs fusion after the coding layers of each model, achieving an accuracy of 86%; Architecture C (late fusion) adopted in this invention performs adaptive gating fusion after each model makes an independent prediction, achieving an accuracy of 91%. The advantage of late fusion is that each model can be optimized independently, the fusion weights can be dynamically adjusted, and the failure of a single model does not affect the overall prediction.

[0079] Predicting performance time series pairs, for example Figure 4 As shown in the figure above, the prediction curves of different models are compared at 100 time steps. The fusion model of this invention (solid line) has the highest degree of fit with the true value (black line). The figure below shows the distribution of prediction error. The mean error of the fusion model of this invention is 0.02 and the standard deviation is 1.45, which is significantly better than the single model (LSTM standard deviation 3.89, Random Forest standard deviation 4.62, XGBoost standard deviation 4.15).

[0080] Experimental results show that the prediction accuracy of the multi-model fusion method of this invention is significantly better than that of a single model and traditional methods. The prediction error of desiccant saturation is less than 5%, the prediction error of remaining service life is less than 2 days, and the F1 score reaches more than 0.94. It can accurately predict and evaluate the status of dehumidifiers, meet the intelligent management needs of maintenance-free dehumidifiers, and has strong environmental adaptability and dynamic optimization capabilities. It can be widely used in the status monitoring scenarios of maintenance-free dehumidifiers for various power equipment.

[0081] This invention integrates three models: LSTM, Random Forest, and XGBoost, fully leveraging the strengths of each: the LSTM model excels at capturing temporal features and can accurately capture the long-term trend of desiccant saturation; the Random Forest model excels at handling nonlinear relationships and can effectively uncover the impact of environmental features on the dehumidifier's state; the XGBoost model has strong generalization ability and can accurately handle the mapping relationship between operating state features and fault risks. Through the synergistic effect of multiple models, prediction accuracy is improved, prediction errors are minimized, and the high-precision prediction requirements of maintenance-free dehumidifiers are met. This invention extracts multi-dimensional features—temporal, statistical, and environmental—through feature engineering, comprehensively covering various factors affecting the dehumidifier's state and avoiding prediction bias caused by single features. The adaptive fusion mechanism dynamically adjusts the weight coefficients based on the real-time prediction performance of each sub-model using an exponentially weighted moving average method, ensuring that the prediction results are always optimal. Simultaneously, through online update mechanisms and retraining trigger conditions, the model can adapt to complex and changing power operating environments, avoiding model performance degradation and maintaining stable prediction results over the long term. This invention can not only accurately predict desiccant saturation, but also predict remaining service life and classify fault risk levels, achieving saturation trend prediction 7-14 days in advance. This facilitates maintenance personnel to formulate maintenance plans in advance, truly realizing intelligent management of maintenance-free dehumidifiers. At the same time, the prediction results of the three sub-models can be analyzed independently, which can clearly locate the key factors affecting the state of the dehumidifier, facilitate fault diagnosis and model optimization, improve the practicality and operability of the method, reduce maintenance costs, and ensure the safe and stable operation of transformers.

[0082] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for predicting the condition of a maintenance-free dehumidifier based on multi-model fusion, characterized in that, Includes the following steps: S1: Data Acquisition: Real-time acquisition of dehumidifier operation data through multi-source sensors, including humidity data, temperature data, pressure data, and vibration data; S2: Data preprocessing: The collected raw data is cleaned, outlier handling is performed, missing value imputation is performed, and data normalization is performed to obtain a standardized dataset; S3: Feature Engineering: Extracting temporal, statistical, and environmental features from standardized datasets to construct multi-dimensional feature vectors; S4: Multi-model parallel prediction: The feature vectors are input into the LSTM model, the random forest model and the XGBoost model respectively for parallel prediction, and the prediction results of the three sub-models are obtained. S5: Adaptive Fusion: Based on the prediction performance of each sub-model on historical data, dynamically calculate the weight coefficients, weight and fuse the prediction results of the three sub-models, and output the final prediction result. S6: Condition Assessment: Based on the fusion prediction results, assess the current condition of the desiccant, predict the desiccant saturation and remaining service life.

2. The method for predicting the condition of a maintenance-free dehumidifier based on multi-model fusion according to claim 1, characterized in that: In step S2, outlier handling is performed during data preprocessing. The criteria identify and replace outliers, and the range of outlier determination is: , The mean of the data. The standard deviation of the data is given; missing values ​​are filled using linear interpolation, and when more than 3 consecutive sampling points are missing, the mean of the previous and next 10 sampling points is used for filling; data normalization uses the Min-Max method, mapping each feature data to the [0,1] interval, and the normalization formula is: ， in, The normalized data has a value range of [0,1] and is used to eliminate the influence of different dimensions of data to make the feature data comparable; x is the original data to be normalized, that is, a single feature data after data collection, cleaning, outlier processing, and missing value imputation. It is the minimum value in the dataset for this feature, that is, the minimum value among all the original data for the same feature; It is the maximum value in the dataset for this feature, that is, the maximum value among all the original data for the same feature.

3. The method for predicting the condition of a maintenance-free dehumidifier based on multi-model fusion according to claim 1, characterized in that: In S3, the time series features include sliding window mean, sliding window variance, trend slope, and periodic components. The mean, variance, and standard deviation are calculated using a 24-hour sliding window, the trend slope is calculated using the least squares method, and the periodic components are extracted using Fourier transform. Statistical characteristics include maximum, minimum, median, quartiles, skewness, and kurtosis. Calculate the above statistical characteristics for each data point over the past 7 days. Environmental characteristics include ambient temperature, ambient humidity, atmospheric pressure, and seasonal factors, with the seasonal factors corresponding to 1 / 2 / 3 / 4 for spring / summer / autumn / winter respectively.

4. The method for predicting the condition of a maintenance-free dehumidifier based on multi-model fusion according to claim 1, characterized in that: In S4, the LSTM model is used to process temporal features. Its model structure includes an input layer, two LSTM hidden layers, a fully connected layer, and an output layer. The input time window is set to 24-72 hours, the number of neurons in the hidden layer is 64-128, the dropout rate is controlled at 0.2-0.3, the model training uses the Adam optimizer with a learning rate of 0.001, the number of training epochs is 100, an early stopping mechanism is used to prevent overfitting, and the output is the predicted value of desiccant saturation for the next 24-72 hours.

5. The method for predicting the condition of a maintenance-free dehumidifier based on multi-model fusion according to claim 4, characterized in that: In S4, the random forest model is used to process environmental features. The number of decision trees is set to 100-200, the maximum depth is 10-20 layers, the minimum number of split samples is 5-10, the feature sampling ratio is 0.6-0.8, and the Gini coefficient is used as the splitting criterion. It mainly explores the nonlinear relationship between environmental features and desiccant saturation and outputs the predicted value of desiccant saturation.

6. The method for predicting the condition of a maintenance-free dehumidifier based on multi-model fusion according to claim 5, characterized in that: In S4, the XGBoost model is used to process the running state features. Its learning rate is 0.01-0.1, the number of trees is 100-300, the maximum depth is 3-8 layers, the L2 regularization coefficient is 0.1-1.0, and the mean squared error (MSE) is used as the loss function. It mainly captures the mapping relationship between running state features and fault risk, and outputs fault risk-related predicted values.

7. The method for predicting the condition of a maintenance-free dehumidifier based on multi-model fusion according to claim 1, characterized in that: In S5, the adaptive fusion weight coefficient calculation method is as follows: based on the root mean square error (RMSE) of each sub-model over the most recent N time windows, the weights are dynamically adjusted using an exponentially weighted moving average method, where N is a positive integer. The weight calculation formula is as follows: ， The fusion calculation formula is as follows: ； in, is the weight coefficient of the i-th sub-model, where i takes the values ​​1, 2, and 3, corresponding to the LSTM model, random forest model, and XGBoost model, respectively; exp is the natural exponential function; This is an adjustment coefficient used to adjust the degree of influence of prediction error on the weighting coefficient, with a value of 2.

0. is the root mean square error of the i-th sub-model over the most recent N time windows, used to measure the prediction accuracy of the i-th sub-model. The smaller the error, the larger the weight coefficient. The summation symbol is 'j', which takes values ​​of 1, 2, and 3, corresponding to the three sub-models. It is used to sum the exponential terms of the three sub-models. This is the final prediction result after multi-model fusion; These are the weight coefficients of the LSTM model; These are the independent prediction results of the LSTM model; These are the weight coefficients of the random forest model; These are the independent prediction results from the random forest model; These are the weight coefficients of the XGBoost model; These are the independent prediction results from the XGBoost model.

8. The method for predicting the condition of a maintenance-free dehumidifier based on multi-model fusion according to claim 1, characterized in that: In S6, the status assessment includes: desiccant saturation assessment, which divides saturation into four levels: 0-30% is normal, 30%≤saturation<60% is slightly saturated, 60%≤saturation<80% is moderately saturated, and saturation≥80% is heavily saturated; remaining service life prediction, which predicts the number of days required for the desiccant to reach the 80% saturation threshold; and failure risk classification, where remaining service life>14 days is low risk, 7 days<remaining service life≤14 days is medium risk, and remaining service life≤7 days is high risk.

9. The method for predicting the condition of a maintenance-free dehumidifier based on multi-model fusion according to claim 1, characterized in that: It also includes an online model update step: every 7 days, the three sub-models and the fusion strategy are incrementally trained using the latest collected running data, and the model parameters are updated; the prediction error of each sub-model is monitored in real time, and if the average prediction error of a sub-model exceeds a preset threshold for 30 consecutive days, model retraining is triggered. The model training adopts a transfer learning strategy, first pre-training with a publicly available power equipment operation dataset, and then fine-tuning with field data.

10. A maintenance-free desiccant condition prediction system based on multi-model fusion, applied to the maintenance-free desiccant condition prediction method based on multi-model fusion described in any one of claims 1-9, characterized in that: The system includes a data acquisition module, a data preprocessing module, a feature engineering module, a multi-model prediction module, a fusion decision module, and a state assessment module. The data acquisition module is used to collect humidity, temperature, pressure, and vibration data related to the operation of the dehumidifier in real time through multi-source sensors. The data preprocessing module is used to clean the collected raw data, handle outliers, imputate missing values, and normalize the data to output a standardized dataset. The feature engineering module is used to extract time-series features, statistical features, and environmental features from the standardized dataset to construct a multi-dimensional feature vector. The multi-model prediction module incorporates LSTM, random forest, and XGBoost models to perform parallel predictions on feature vectors and output independent prediction results for each sub-model. The fusion decision module dynamically calculates weight coefficients based on the historical prediction performance of each sub-model and performs weighted fusion of the prediction results to output the final prediction result. The state assessment module assesses the current operating state of the dehumidifier, predicts the desiccant saturation and remaining service life, and classifies the failure risk level based on the final prediction result.

Images