Temporal deep learning model predicts aging of power composite insulators

Power transmission networks depend on composite insulators to keep high-voltage conductors safely separated from the ground, yet these components age continuously under ultraviolet radiation, temperature swings, moisture, surface pollution, corona discharge and constant electrical stress. As the materials deteriorate, they lose hydrophobicity, develop surface erosion and internal microcracks, and their leakage current rises, raising the probability of flashover, insulation failure and unexpected grid shutdowns. A study published in Discover Artificial Intelligence on 27 October 2026 now reports a temporal deep learning framework, called ModABi-LSTMO, that the authors say can classify the discrete aging states of power hard composite insulators with 98.9 percent accuracy on a publicly available benchmark dataset.

The framework, developed by Tongbo He, Xuexing Dong, Xianzhong E, Weiwei Yang, Shizeng Liu and colleagues, with Jialun Wan among the co-authors, combines three building blocks. Discrete Wavelet Transform (DWT) extracts time-frequency features from leakage current signals, a Bidirectional Long Short-Term Memory (BiLSTM) network learns how degradation unfolds forward and backward through time, and a Modified Archimedes Optimization (ModAO) strategy tunes the network’s parameters to improve convergence and stability. According to the paper, this integration is intended to overcome the weaknesses of earlier approaches, which the authors describe as limited in prediction accuracy, adaptability and temporal feature analysis.

The data behind the study is the Power Hard Composite Insulator Aging Dataset from Kaggle, containing 9,775 timestamped records of simulated operational and environmental sensor observations designed to emulate the aging behavior of high-voltage composite insulators under representative operating conditions. The authors are explicit that this is a simulated smart-grid monitoring testbed rather than data from a known utility transmission grid, and that sensor hardware specifications, geographical deployment, sampling instrumentation and labeling processes cannot be independently verified. They therefore treat the dataset as a benchmark for algorithm development and comparison, not as a substitute for field measurements.

Preprocessing began with time-series interpolation to fill missing values in electrical load, temperature, humidity, contamination index and leakage current readings, using forward-backward and linear filling within sliding windows to preserve sequence continuity. Min-Max normalization then scaled all sensor features to a 0-to-1 range so that large-valued variables would not dominate model training. The leakage current signals were decomposed with the Daubechies-4 (db4) wavelet basis over four levels, chosen for its strong localization properties in non-stationary electrical signals. Low-pass filtering produced approximation coefficients representing slow, long-term degradation trends, while high-pass filtering produced detail coefficients capturing abrupt changes, discharge activity and abnormal insulation conditions.

The BiLSTM layer then processes these wavelet features in both temporal directions, concatenating forward and backward hidden states so the model can represent short-term fluctuations and long-term degradation patterns simultaneously. On top of this, ModAO performs offline hyperparameter optimization, searching over learning rate, number of hidden units, batch size, dropout rate and L2 regularization coefficient, with validation F1-score as the objective function. Each candidate configuration was trained for a maximum of 20 epochs with early stopping, using a population of 20 candidate solutions over 50 iterations. Once the best hyperparameters were found, the final BiLSTM was retrained over a full 100-epoch schedule with the Archimedes Optimization Algorithm handling weight updates. The task itself is multi-class classification of a target variable, Aging_State, covering aging stages I through VI.

The reported results are consistent with the headline figure. The model achieved 98.9 percent accuracy, 98.2 percent precision, 98.7 percent recall and 98.5 percent F1-score, outperforming re-implemented baselines including CNN-LSTM at 96.1 percent, BiLSTM alone at 95.8 percent, and an ensemble-based framework at 98.5 percent, with corresponding gaps in precision and recall. An ablation analysis traced the contribution of each component: a BiLSTM with AOA-based optimization reached 97.6 percent accuracy, standard AOA-based hyperparameter selection raised this to 98.1 percent, the proposed ModAO lifted it to 98.4 percent, and the full ModABi-LSTMO pipeline reached 98.9 percent, indicating that the gains come from the optimization strategy as well as the architecture.

Robustness checks reinforced the picture. Five-fold stratified cross-validation produced standard deviations of 0.21 for accuracy, 0.17 for precision, 0.21 for recall and 0.20 for F1-score, and paired t-tests across the five folds yielded p-values below 0.001 for all performance measures. Error analysis showed the proposed model had the lowest variance, 0.0421, compared with 0.0485 for ModAO alone and 0.0820 for DWT alone, and the smallest mean absolute error at 0.0254. Interpretability was addressed through SHAP analysis, which ranked leakage current as the dominant contributor to predictions with an importance of 0.480, followed by humidity at 0.310 and pollution index at 0.270, reflecting mechanisms in which moisture and contamination raise surface conductivity and create leakage pathways.

The authors also report deployment-relevant timings. After offline training, inference takes roughly 12.8 milliseconds per sample on an NVIDIA RTX 2060 GPU and 47.6 milliseconds per sample on an Intel Core i5 CPU, which they argue is fast enough for periodic smart-grid monitoring and edge-assisted monitoring devices with constrained computing power. In their framing, the framework supports continuous condition monitoring and predictive maintenance scheduling by estimating deterioration trends, reducing reliance on manual inspection, visual assessment and conventional diagnostic criteria that are labor-intensive, subjective and prone to missing early-stage degradation.

The study’s limitations are stated plainly. The authors note that the relatively small sample size may affect generalization across operating conditions and environments, and that the model’s results may vary with the dynamics or severity of the environment. Because the dataset is simulated, it cannot fully reflect real-world complexities such as degradation mechanisms, climatic extremes, geographical variation and installation differences. The paper states that practical deployment would require validation with separate large-scale utility field data and laboratory aging experiments, and that future work will incorporate measurements from utilities, laboratory aging tests and multiple locations to improve feature diversity, model generalization and applicability to real transmission systems.

Within those boundaries, the study offers a concrete template for temporal machine learning applied to insulator health: wavelet-based time-frequency features feeding a bidirectional recurrent network, tuned by a metaheuristic optimizer and validated with cross-validation, statistical testing and interpretability analysis. Whether the reported 98.9 percent accuracy transfers from a simulated benchmark to live transmission infrastructure remains the central open question, and the authors themselves position their model as a demonstration of predictive intelligence under representative conditions rather than a replacement for field validation. The work is open access, published in Volume 6, article number 1668 of Discover Artificial Intelligence.

Subject of Research: Temporal deep learning for aging assessment of power hard composite insulators

Article Title: Construction method of temporal deep learning model for aging assessment of power hard composite insulators

Article References: He, T., Dong, X., E, X., Yang, W., Liu, S., & Wan, J. (2026). Construction method of temporal deep learning model for aging assessment of power hard composite insulators. Discover Artificial Intelligence, 6(1), Article 1668. https://doi.org/10.1007/s44163-026-02486-4

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02486-4

Keywords: composite insulators, deep learning, BiLSTM, discrete wavelet transform, Archimedes optimization, predictive maintenance, leakage current, power transmission, smart grid, SHAP interpretability, simulated benchmark dataset, insulator aging