Residual Mixture-of-Experts LSTM Models for Multi-Regional Wind Speed Forecasting

Wind speed forecasting is vital for energy planning and grid integration in regions with high wind potential and growing renewable energy deployment. Conventional recurrent models such as Long Short-Term Memory (LSTM) networks capture temporal dependencies effectively but struggle to generalize across heterogeneous wind regimes. To overcome these limitations, this study proposes a Residual Mixture-of-Experts (ResMoE) architecture, building on the Mixture-of-Experts (MoE) framework, where multiple LSTM experts are combined through a gating mechanism to model different components of the data. The ResMoE introduces a global expert that captures large-scale temporal dynamics within each province, along with residual experts that model localized deviations as additive corrections. This decomposition enables the model to separate shared patterns from regime-specific variations, improving both learning stability and specialization. Experiments were conducted using long-term meteorological datasets (~210,000 hourly records per site) from five representative provinces in Turkey (Balikesir, Çanakkale, İzmir, Kirklareli, and Tekirdaǧ) and each province is characterized by distinct wind regimes. Input features included 9 meteorological variables, with a 24-hour lag window used to model short-term dependencies. Models were trained using the AdamW optimizer with learning rate scheduling and early stopping, and evaluated using the Mean Absolute Percentage Error (MAPE) across multiple randomized runs, along with complementary error metrics and supported by statistical significance analysis. Results show that the proposed ResMoE consistently achieves the lowest forecasting errors across all provinces. Across the five study sites, ResMoE reduces MAPE by approximately 6.5-10.3% relative to the baseline LSTM and by about 1.8-6.0% relative to the standard Mixture-of-Experts (MoE). The residual formulation improves training stability by anchoring predictions to a global baseline and reduces expert collapse by limiting each expert to modeling residual errors. Additionally, the additive structure allows direct analysis of expert contributions, providing a more interpretable forecasting framework. Overall, ResMoE provides a structured and robust approach for heterogeneous time-series forecasting, with applicability beyond wind energy.