TCN+LightGBM: A Hybrid Deep Learning–Gradient Boosting Architecture for Interpretable Short-Term Electricity Demand Forecasting in Bangladesh

Main contribution:
The paper proposes a novel two-stage hybrid architecture for short-term electricity demand forecasting on Bangladesh’s PGCB grid data that is structurally different from prior hybrid approaches. Instead of fusing two neural branches end-to-end (as in [1] and [2]) or combining two independently-trained model outputs via linear regression (as in Gong et al. [4]), this study uses an already-trained TCN purely as a feature extractor: it takes the 32-dimensional learned embedding from the TCN’s penultimate layer and concatenates it with 49 hand-engineered tabular features (weather, calendar, economic, lag/rolling statistics) to form an 81-dimensional input, on which a LightGBM regressor is trained to produce the final prediction.

Why this matters (the novelty claim):

The final predictor is a tree ensemble, not a black-box neural network — so exact, native feature-importance analysis is possible instead of an approximate/surrogate explainer, which is a genuine interpretability gain over prior Bangladesh-grid hybrids.
It achieves 97.26% accuracy (2.74% MAPE) on the full hold-out set, and 1.87% MAPE on a literature-matched Jan–Feb 2024 window — beating Islam et al.’s 2.30% MAPE on the same window, and substantially better than Bhadra & Rois’s 5.45% MAPE (though that comparison is caveated by daily vs. hourly granularity).
The study is methodologically thorough: it adds a five-part evaluation protocol (point accuracy, statistical significance via Diebold-Mariano test, exact feature-importance/explainability, quantile-based uncertainty coverage, multi-step 24-hour forecasting, and 5-fold walk-forward robustness validation) that neither reference paper applies in full.
It also honestly reports a nuance/limitation rather than overselling: the Diebold-Mariano test shows standalone TCN actually has significantly lower squared error than the hybrid — meaning the hybrid’s advantage is specifically in relative-percentage error (MAPE), not absolute-MW error, a trade-off the paper explicitly flags for grid operators (peak-hour reserve planning vs. day-ahead trading use cases).