Multi-Horizon Transformer Oil-Temperature Forecasting: Temporal Dependence, Load-Variable Utility, and Model Complexity

Abstract

Accurate transformer oil-temperature forecasting is important for thermal-risk assessment and operational planning. However, reported gains from complex forecasting models may be affected by future information leakage, weak seasonal baselines, inconsistent target periods, and test-based model selection. This study establishes a leakage-free, target-aligned framework for direct forecasting at 6, 12, and 24 h, integrating controlled model comparison, input-utility analysis, exact temporal interpretation, and cross-dataset confirmation. Only information available at or before the forecast origin is used, and identical validation and test target periods are maintained across horizons and lookback lengths. Naive predictors, regularized autoregression, ensemble methods, XGBoost variants, deep sequence models, and linear-nonlinear hybrids are evaluated using expanding-window validation and moving-block bootstrap analysis. OT-only Ridge regression with a 72 h lookback and $\lambda = 10^{-4}$ was selected for all three horizons, achieving ETTh2 RMSEs of 3.1568, 4.0042, and 4.2920. After ETTh1-specific refitting, the corresponding RMSEs were 1.3701, 1.7532, and 2.0976. Ridge significantly outperformed the daily-seasonal baseline at 6 and 12 h, while the 24 h gain was not statistically distinguishishable. The six historical load channels provided no robust incremental value. Exact Ridge contributions showed a shift from recent thermal persistence at 6 h to dominant daily-cycle dependence at 24 h. Rapid cooling was overpredicted and rapid heating was underpredicted. Overall, increased model complexity did not provide a consistent advantage under a controlled protocol, while the combined evaluation, interpretation, and cross-dataset confirmation offer reproducible empirical guidance for transformer oil-temperature forecasting.