LSTM Networks with Custom Loss Function for California Energy Price Prediction
A recent study available on arXiv introduces an innovative online learning framework aimed at predicting next-day electricity prices in California's energy market. This system employs Long Short-Term Memory (LSTM) networks, factoring in historical pricing, weather patterns, and energy generation methods. Researchers developed a unique custom loss function that integrates Mean Absolute Error, Jensen-Shannon Divergence, and a smoothness penalty to bolster accuracy and interpretability. The findings demonstrate that this tailored approach significantly enhances the model's predictive performance, aligning forecasts with actual market prices, a vital improvement for grid operators, energy suppliers, and consumers.
Key facts
- Study focuses on day-ahead electricity price prediction in California.
- Uses LSTM networks for forecasting.
- Features include historical price data, weather conditions, and energy generation mix.
- Custom loss function integrates MAE, JSD, and smoothness penalty.
- Adaptive online learning framework allows incremental model updates.
- Results show improved prediction accuracy with the custom loss function.
- Published on arXiv with ID 2510.16898.
- Relevant to grid operators, energy producers, and consumers.
Entities
Institutions
- arXiv
Locations
- California