Accurate forecasting of residential electricity demand is increasingly important as modern power systems must balance variable renewable generation with fluctuating household loads. A new comparative study evaluates two advanced deep learning architectures—LSTM networks with post-hoc explanations and Temporal Fusion Transformers (TFT)—for short-term appliance load prediction.
The research addresses a fundamental challenge: household electricity consumption exhibits high variability, sudden spikes, and non-stationary behavior that confounds traditional statistical methods. The study implements both an MSE-trained LSTM explained through SHAP values and an MSE-trained TFT with built-in attention mechanisms, evaluated on identical datasets and feature-engineering procedures to ensure fair comparison.
Results demonstrate that the TFT achieves measurably lower forecasting errors than the LSTM, with superior performance on standard metrics including RMSE, MAPE, and R². Beyond point estimates, the quantile-trained TFT also provides probabilistic forecasts with well-calibrated uncertainty bounds—achieving 81.5% coverage probability on 80% prediction intervals, indicating reliable confidence estimation.
A critical advantage of the TFT approach is its inherent interpretability. The model's variable-selection and attention mechanisms reveal which features and time steps most influence predictions, providing feature-level and temporal explanations without requiring post-hoc analysis. The LSTM-SHAP combination offers complementary explanations but requires additional computational overhead.
For grid operators and utility planners, these advances enable more reliable demand forecasting to support real-time balancing, capacity planning, and integration of distributed resources. The study demonstrates that modern deep learning models need not sacrifice interpretability for accuracy—a critical requirement for deployment in regulated energy environments where model decisions must be auditable and explainable to stakeholders and regulators.
