Improving multi-horizon LSTM-based electricity price forecasting through outlier handling and advanced optimization

Abstract Electricity price forecasting is crucial for market participants seeking to reduce risk and improve bidding strategies. This paper examines the combined impact of data preprocessing and hyperparameter optimization on Long Short-Term Memory (LSTM) models for forecasting the Italian day-ahead electricity price across the hour-, day-, and week-ahead horizons. Two outlier handling methodologies, namely dynamic thresholding and winsorization, are applied, and their impacts on key statistical properties of the training dataset are compared. The LSTM model is optimized by adopting Grey Wolf Optimization, Dung Beetle Optimization, and the MSADBO algorithm. A Sequence-to-Sequence architecture is thus used to obtain multi-horizon forecasts. Results comparison confirms that the optimized LSTM models achieve the best performance across all horizons, not only the baseline LSTM but also established machine learning benchmarks. These findings highlight the importance of statistically informed preprocessing and advanced optimization algorithms for improving the accuracy of multi-horizon electricity price forecasting.