A Hybrid Temporal-spectral Load Forecasting Model with Static Context Fusion for Smart Cities

Accurate short-term electricity load forecasting is essential for reliable and efficient smart city energy management, particularly in environments characterized by high-dimensional, heterogeneous, and noisy multivariate signals. However, existing forecasting models often struggle to simultaneously capture nonlinear temporal dependencies, multi-scale periodicity, and static contextual influences within a unified framework. To address this challenge, this study proposes a hybrid deep learning architecture that integrates Bidirectional Long Short-Term Memory (BiLSTM) for temporal modeling, an additive attention mechanism for adaptive time-step weighting, Fast Fourier Transform (FFT)-based frequency residual learning for periodicity extraction, and embedding-based static feature fusion for contextual representation. The model is evaluated on the ISO-NE Smart City Energy Dataset for next-hour electricity load forecasting using a two-week input window (336 hours). Experimental results demonstrate that the proposed hybrid framework significantly improves predictive accuracy, achieving an RMSE of 25.51 kW and an R$$^2$$ of 0.9905, outperforming recurrent, convolutional, and transformer-based baselines under identical evaluation settings. Ablation analysis confirms that temporal attention and frequency-domain residual modeling contribute substantially to performance gains. These findings indicate that joint temporal–spectral modeling combined with static contextual fusion provides a robust and effective solution for complex smart-city electricity forecasting tasks.