Time-Series Forecasting

A foundational approach where future dissolved gas levels are predicted based on a linear combination of their own past values. ARIMA and SARIMA models capture trend and seasonality in gas concentration trajectories.

• Lag Order (p): Defines how many past time steps influence the next prediction
• Differencing (d): Applied to remove non-stationarity in trending gas accumulation data
• Moving Average (q): Models the dependency between an observation and residual errors from a moving average of prior values

Example: An ARIMA(2,1,1) model predicts acetylene levels using the two prior readings, first-order differencing, and one lagged forecast error.

LSTM architectures are a class of recurrent neural networks engineered to overcome the vanishing gradient problem, enabling them to learn dependencies spanning hundreds of time steps in sensor data.

• Forget Gate: Determines which information to discard from the cell state, allowing the network to reset when transformer operating modes shift
• Input Gate: Controls which new gas concentration values update the memory cell
• Output Gate: Regulates what information from the cell state is exposed to the next layer

LSTMs excel at modeling the non-linear degradation trajectories of transformer insulation where failure precursors may appear weeks before an event.

TFT is an attention-based architecture purpose-built for interpretable multi-horizon forecasting, combining recurrent layers with self-attention mechanisms.

• Variable Selection Networks: Automatically identify which input features—such as load current, ambient temperature, or specific gas ratios—are most relevant at each time step
• Gated Residual Networks: Enable skip connections that allow the model to adapt its depth dynamically, preventing overfitting on sparse DGA data
• Quantile Regression Output: Produces prediction intervals (e.g., P10, P50, P90) rather than point estimates, critical for risk-based maintenance scheduling

TFT provides native explainability by exposing attention weights that show which historical periods most influenced a fault forecast.

Effective transformer forecasting models incorporate exogenous variables—external factors that influence gas generation rates but are not themselves predicted by the model.

• Load Current: Higher loading increases winding temperature, accelerating cellulose degradation and gas production
• Ambient Temperature: Seasonal variations affect cooling efficiency and oil circulation patterns
• Tap Changer Operations: Frequent voltage adjustments generate transient arcing that elevates acetylene levels
• Moisture Content: Water accelerates hydrolysis, producing carbon oxides at rates unrelated to electrical stress

Multivariate models that ignore these covariates often generate false alarms during routine load peaks.

Real-world DGA sensor data rarely arrives at uniform intervals due to communication dropouts, sensor maintenance windows, or manual sampling schedules. Specialized techniques address this irregularity.

• Time-Aware Embeddings: Encode the actual timestamp delta between observations directly into the model input, rather than assuming fixed spacing
• Neural Ordinary Differential Equations: Model gas concentration dynamics as a continuous-time process, evaluating the hidden state at arbitrary query points
• Imputation with Uncertainty: Use Gaussian processes to estimate missing values while propagating the uncertainty of those estimates through the forecast

Ignoring irregular sampling can introduce systematic bias, as manual samples are often taken precisely when operators suspect a problem.

Point forecasts of future gas levels are insufficient for risk management. Probabilistic forecasting quantifies uncertainty by outputting full predictive distributions.

• Prediction Intervals: A 90% interval for acetylene at 30 days ahead enables setting alarm thresholds that balance false positives against missed faults
• Conformal Prediction: A distribution-free framework that wraps any point forecaster to produce statistically valid uncertainty estimates without distributional assumptions
• DeepAR: An autoregressive recurrent network trained to output parametric distributions (e.g., Gaussian or negative binomial) at each horizon step

Asset managers use probabilistic outputs to calculate the probability of exceeding IEC 60599 normal limits within a given planning window.

LSTM and Temporal Fusion Transformer models ingest years of online DGA monitor data to forecast the concentration trajectory of fault gases like acetylene and hydrogen 30 to 90 days into the future. By establishing a dynamic baseline for each transformer, these models detect subtle rate-of-change anomalies that static threshold alarms miss. A steadily rising ethylene trend, even within acceptable limits, can signal developing thermal fault conditions in cellulose insulation, triggering a proactive oil sampling recommendation before a Duval Triangle classification becomes critical.

Predictive models combine historical load current, ambient temperature, and cooling system status time-series data to forecast winding hot-spot temperature hours ahead of real-time. This application of IEEE C57.91 thermal models within a deep learning framework allows operators to anticipate dangerous temperature excursions before they accelerate Degree of Polymerization (DP) degradation. The forecast enables preemptive load shedding or cooling bank activation, directly extending the Remaining Useful Life (RUL) of aging transformer fleets.

Time-series forecasting tracks the migration of moisture content between solid insulation and insulating oil as a function of load-driven temperature cycles. By predicting future water content in oil, these models identify the optimal window for online degassing or oil reclamation before free water formation compromises dielectric strength. The approach integrates Karl Fischer titration lab results with continuous relative saturation sensor data to model the complex diffusion dynamics governed by Arrhenius reaction kinetics.

Load Tap Changer (LTC) diagnostics leverage time-series analysis of motor torque signatures, contact temperature, and dissolved gas spikes during tap change operations. Forecasting models trained on vibration signature data and arcing duration trends predict mechanical wear progression, identifying the optimal maintenance interval before a catastrophic LTC failure occurs. This is critical because LTC failures represent the single most common cause of major transformer outages, and forecasting avoids both premature maintenance and unexpected breakdowns.

Multiple time-series forecasts from DGA, tan delta, furan analysis, and partial discharge monitors are fused into a unified Health Index trajectory using ensemble learning techniques. A Random Forest or XGBoost meta-model weights each diagnostic stream based on its historical predictive power for specific transformer designs. The resulting composite score forecasts the rate of health degradation, enabling Condition-Based Maintenance (CBM) scheduling that prioritizes assets approaching a critical threshold months before traditional time-based maintenance would intervene.

Long-term time-series models incorporate sensor drift compensation algorithms to maintain forecast accuracy as online DGA monitors age. By modeling the gradual degradation of sensor calibration as a latent variable within a state-space model, the system distinguishes true gas trends from measurement artifacts. This prevents false alarms caused by sensor degradation while preserving sensitivity to genuine fault signatures, ensuring that IEC 60599 compliance is maintained without requiring manual recalibration of every field device.

A specialized recurrent neural network architecture designed to learn long-term dependencies in sequential data. LSTMs use gating mechanisms—input, forget, and output gates—to control information flow, mitigating the vanishing gradient problem. In transformer diagnostics, LSTMs process historical dissolved gas analysis (DGA) readings to predict future gas concentration trajectories, capturing the temporal dynamics of fault evolution over weeks or months.

An attention-based architecture purpose-built for interpretable multi-horizon forecasting. TFT combines recurrent layers with self-attention to handle static metadata, known future inputs, and observed historical inputs simultaneously. Key advantages for predictive maintenance:

• Variable selection networks identify which gas sensors are most relevant
• Quantile regression outputs prediction intervals, not just point forecasts
• Interpretable attention reveals which time steps the model weights most heavily

The process of transforming raw sensor time-series into predictive inputs. Critical engineered features include:

• Gas ratios: C2H2/C2H4, CH4/H2, C2H4/C2H6 per IEC 60599
• Rolling statistics: 7-day and 30-day moving averages of hydrogen concentration
• Rate-of-change: First and second derivatives of acetylene levels
• Lag features: Gas values from t-1, t-7, t-30 days
• Cyclical encoding: Sin/cos transforms of seasonal load patterns

AutoRegressive Integrated Moving Average models serve as essential baseline benchmarks before deploying deep learning. ARIMA captures linear trends and seasonality in univariate gas concentration series. In transformer monitoring, a well-tuned ARIMA model often outperforms complex neural networks on short-horizon forecasts with limited data, making it a critical sanity check. The Box-Jenkins methodology guides identification of p, d, q parameters through ACF and PACF analysis.

Moving beyond point predictions to output full probability distributions of future gas levels. Techniques include:

• Quantile regression: Predicts the 10th, 50th, and 90th percentiles
• DeepAR: Amazon's autoregressive RNN that outputs parametric distributions
• Conformal prediction: Distribution-free uncertainty quantification with guaranteed coverage This enables risk-based maintenance scheduling—acting when the upper bound of acetylene exceeds a threshold, not just the mean.

A critical design choice in transformer time-series modeling. Univariate models forecast a single gas (e.g., hydrogen) using only its own history—simpler and robust to missing data. Multivariate models leverage cross-correlations between gases, load current, ambient temperature, and oil temperature simultaneously. While multivariate approaches capture coupled fault signatures, they require careful handling of sensor drift and asynchronous sampling rates across different monitoring systems.