A Spike-Resilient Temporal-Adaptive Neural Framework for Day-Ahead Electricity Price Interval Forecasting

Electricity price data demonstrate spike-induced heteroscedasticity and non-stationary market dynamics. To tackle these challenges, this paper proposes a spike-resilient temporal-adaptive neural framework for day-ahead electricity price interval forecasting. The framework integrates a robust Box-Cox transformation with Huber loss to mitigate outliers and heavy tails. It optimizes the parameters of the robust Box-Cox transformation via gradient descent instead of Newton’s method for improved computational efficiency. Additionally, a dynamic weighted moving average algorithm is embedded to construct prediction intervals, adaptively capturing time-varying patterns. Experimental results confirm that the proposed approach outperforms existing benchmarks in electricity price interval forecasting, enhancing predictive reliability and sharpness.