TPA-net: Temporal pyramid convolution network with attentional BiLSTM for power load forecasting

Accurate short-term and long-term power load forecasting plays a pivotal role in ensuring the reliability and economic efficiency of modern smart grids. To address the challenges of complex temporal patterns, multi-scale periodicity, and dynamic variability in load data, we propose a novel deep forecasting framework named TPA-Net (Temporal Pyramid CNN–BiLSTM–Attention). The proposed architecture consists of three key components: a temporal pyramid multi-scale convolutional module to extract hierarchical periodic features across hourly, daily, and weekly levels; a bidirectional LSTM (BiLSTM) module to capture global temporal de-pendencies in both forward and backward directions; and an attention mechanism to dynamically emphasize critical time steps. Extensive experiments conducted on real-world power consumption datasets demonstrate that TPA-Net consistently outperforms state-of-the-art baselines across multiple forecasting horizons (1-h, 6-h, and 24-h), achieving significant improvements in RMSE, MAPE, and R 2 metrics. These results highlight the effectiveness and generalizability of TPA-Net in complex load forecasting scenarios.