Universal Dual-Horizon Residential Forecasting Model Using Trigonometric Deep Long Short-Term Memory with Attention Mechanisms

Effective forecasting in residential energy systems is important for efficient grid operation. However, existing methods face challenges with the rapid growth of data heterogeneity, the intermittent nature of renewables, and accurate electricity price forecasting. Further, many approaches have failed to model a unified framework for forecasting multiple variables simultaneously. To overcome these limitations, we propose a multi-input multi-output universal model using the Trigonometric Deep Long Short-Term Memory with Attention Mechanisms (Trigo-Deep LSTM-AM) for dual-horizon residential forecasting. The proposed model predicts both short-term (day-ahead) and long-term (year-ahead) residential load demand, electricity prices, and solar and wind generation. It integrates LSTM networks, attention mechanisms, and trigonometric encoding to effectively capture periodic and non-periodic patterns. The proposed model was evaluated on the testing dataset (the final 15% of the 2015-2020 UK dataset) and further validated using residential datasets from Germany (2016-2020) and the USA (2017-2021). For the UK testing dataset, the proposed model achieves average MSE-based accuracy improvements of approximately 19%, 53%, 61%, and 34% for load, price, solar, and wind generation compared with RNN-AM, CNN-AM, and GRU-AM baselines. This could enhance resource allocation, grid stability, and cost efficiency in power system management.