A1405
Title: Adaptive forecasting of electricity spot prices: A hybrid HMM-LSTM approach enabling regime detection and prediction
Authors: Christina Erlwein-Sayer - University of Applied Sciences HTW Berlin (Germany) [presenting]
Tilman Sayer - mobile-de (Germany)
Abstract: Electricity spot prices are characterized by frequent volatility and sudden price jumps, which complicates accurate intraday price forecasting. Forecasting models often struggle to adapt to the dynamic nature of these price movements, as market conditions can shift rapidly. The aim is to propose a novel hybrid approach that combines the predictive power of Long Short-Term Memory (LSTM) networks with the flexibility of Hidden Markov Models (HMMs) to better capture these regime shifts. The LSTM architecture is enhanced by incorporating an underlying HMM that detects and adapts to changing market regimes. These regimes are adaptively filtered from the data, enabling the model to separate the spot price series into regime-specific segments. The recurrent neural network is trained for each regime, and the HMM-derived state probabilities are integrated as gates into the prediction model. The integration of HMM with LSTM forecasts enhances both the accuracy and interpretability of the model. Each LSTM forecast is conditioned on the filtered state of the underlying market regime, which allows the model to capture the different price dynamics that occur in varying market conditions. The model is applied to German spot prices, demonstrating its effectiveness in forecasting price movements by leveraging both short- and long-term market dynamics. The HMM-gated LSTM framework outperforms traditional forecasting methods in terms of predictive accuracy.
