A hybrid statistical–machine learning framework for day-ahead electricity price forecasting in smart city environments

Accurate day-ahead electricity price forecasting, particularly in the context of customer-centric services and emerging smart city infrastructures, is a cornerstone of efficient power market operations, supporting bidding strategies, hedging, and grid reliability. This paper proposes SSXLR, a fault-tolerant hybrid forecasting framework that combines Seasonal Autoregressive Integrated Moving Average with Exogenous variables (SARIMAX), Extreme Gradient Boosting (XGBoost), and Long Short-Term Memory (LSTM) within a 2-out-of-3 voting architecture to capture both linear seasonal trends and nonlinear temporal dynamics while reducing dependence on a single forecasting branch. The proposed method was evaluated using one year of historical data (July 1, 2024–July 1, 2025) from Duke Energy Ohio and Kentucky (DEO&K) within the PJM interconnection. The dataset integrates day-ahead market prices ($/MWh), corresponding load consumption (MW), eleven weather variables, and population growth indicators. The target is the day-ahead price profile for July 2, 2025, consisting of 24-hourly points. Results show that SSXLR achieves more than 30% reduction in Root Mean Squared Error over SARIMAX and XGBoost, respectively. The Mean Absolute Error (MAE) reached 1.196, demonstrating the ability of SSXLR to account effectively for external market drivers and temporal dependencies. These findings highlight the scalability of the proposed framework for regional electricity markets, where accurate day-ahead forecasting can improve operational planning, supply–demand balance, and economic efficiency. The proposed framework is dynamic and region-adaptable, allowing it to be transferred to other smart city or regional energy systems when electricity price, load/consumption, weather, and calendar data are available. Beyond improving predictive accuracy, the 2-out-of-3 voting mechanism provides a fault-tolerant forecasting layer by allowing the final output to be determined from the two most consistent forecasting branches when one branch becomes unavailable or produces an unreliable output.