TSD-Net: A Lightweight Trend-Seasonal Decomposition Network for Household Electricity Load Forecasting Under Data-Scarce

Abstract

Short-term household electricity consumption forecasting is vital for demand-side energy management, dynamic pricing, and behind-the-meter storage optimisation. Nonetheless, individual-household datasets often have few daily observations per subject, creating a data-scarce learning problem in which balancing model capacity and generalisation is critical. Recent hybrid deep learning models have integrated attention mechanisms to learn from temporal feature representations, but daily-resolution household prediction remains challenging, especially in small-sample settings. This study provides a controlled comparison of six deep learning models, including LSTM, BiLSTM, BiGRU, a non-decomposed hybrid model, a CBAM-based hybrid model and the proposed lightweight Trend-Seasonal Decomposition Network (TSD-Net). All evaluated models use the same preprocessing procedure, time-based data split, and five-seed evaluation protocol to minimise the effect of random initialisation. Metrics for model complexity and computational cost include trainable parameters, model size, training time, training time per epoch, and test inference time. Among the three competing models, TSD-Net achieves a mean test R² (0.5957), outperforming both the CBAM-based hybrid model (0.5854) and BiGRU (0.5844) on the evaluated household power-consumption dataset. The results show that the explicit trend-seasonal decomposition representation is competitive for household load prediction in data-scarce situations. The results also show that increased architectural complexity does not necessarily improve predictive performance. The TSD-Net model achieves the best accuracy and has considerably fewer parameters than the non-decomposed hybrid. Statistical tests across the five seeds further support whether observed performance differences are reliable across repeated runs. In summary, the proposed study illustrates that a compact decomposition-based architecture delivers an excellent trade-off between prediction performance and model complexity for data-limited household energy-management applications and practical deployment on resource-constrained platforms.