MATSUMOTO Lab - Electricity spot price forecast (en)

Electricity spot price forecast

We update our daily forecasts of electricity prices (JEPX spot prices). Using six different methods (including machine learning algorithms) we generate forecasts and combine them through an advanced online learning ensemble approach.

Based on weekly weather forecasts from the Japan Meteorological Agency, we produce price predictions up to 7 days ahead and evaluate their accuracy.

For related charts and data, including electricity market trading prices, grid information, weather conditions, solar radiation forecasts, and imbalance prices, please refer to the relevant pages.

Data sources:

Japan Electric Power Exchange (JEPX): http://www.jepx.org
Japan Meteorological Agency: https://www.jma.go.jp
OCCTO Wide-area Reserve Margin Disclosure: https://web-kohyo.occto.or.jp/kks-web-public/

Forecast data download (JSON format)

▼ About the Forecasting Model (Click to Expand)

■ Model Overview

We forecast electricity prices at 30-minute intervals using an ensemble approach that combines six different methods: a Generalized Additive Model (GAM), regularized regression (Ridge regression), and four machine learning algorithms: k-nearest neighbors (kNN), artificial neural networks (ANN), support vector regression (SVR), and random forest (RF).

The GAM captures long-term trends through smooth seasonal patterns such as annual cycles, while the other models (Ridge, kNN, ANN, SVR, and RF) are designed to capture short-term fluctuations.

Our ensemble forecasting approach is based on an online machine learning technique called Multivariate Probabilistic CRPS Learning (Berrisch and Ziel, 2023, 2024). Due to the online nature of this learning method, the weights assigned to individual forecasting agents (i.e., models) are updated daily, giving more importance to those that have shown higher accuracy in recent predictions.

For instance, the GAM receives higher weight when long-term trends dominate, while machine learning models take precedence during periods of strong short-term weather sensitivity. This dynamic weighting scheme allows for overall more accurate predictions compared to any single model. Model performance is visualized using cumulative Mean Absolute Error (MAE) comparisons.

For daily average price forecasting, we apply a quantile regression approach, specifically a quantile generalized additive model (qGAM), to generate probabilistic forecasts (i.e., percentile forecasts). Since predictions of higher and lower percentiles are particularly sensitive to the input data, we estimate separate models for short- and long-term variations and combine them to ensure robustness.

The relevant papers are listed below.