Cap-Aware Boosting for Green Day-Ahead Electricity Price Forecasting

Shubham PrataparXiv 2026 · arXiv

TL;DR

This paper introduces a specialized LightGBM-based framework for forecasting 15-minute electricity prices in India's Green Day-Ahead Market (GDAM). By explicitly modeling regulatory price caps through a hybrid classification-regression approach and incorporating domain-specific features like solar curves and auction shapes, the authors achieve a 31.8% reduction in error over state-of-the-art neural time-series models.

What Problem Does This Paper Address?

Forecasting electricity prices in the Indian GDAM is notoriously difficult due to two main factors: high intraday volatility across 96 daily blocks and a structural "ceiling" created by regulatory price caps.

Standard regression models often struggle with these caps; they tend to predict a "smooth" average price, systematically underestimating the frequent events where prices hit the limit (Rs 10,000/MWh). Furthermore, market participants must navigate complex data availability schedules, where information from related markets like the Real-Time Market (RTM) and conventional Day-Ahead Market (DAM) becomes available at different lags. The authors aim to build a system that is both "cap-aware" and "leak-safe," ensuring forecasts only use information available at the time of the auction.

Methods

The core of the proposed system, Prod v1, moves away from a single regressor in favor of a multi-headed LightGBM architecture. As shown in Figure 1, the model splits the task into separate pathways:

  1. Point Forecast Pathway: This uses a binary classifier to estimate the probability that a price will hit the cap ($p_{cap}$) and a log-price L1 regressor to predict the price if it stays below the cap. The final forecast is a probabilistic blend of these two outputs.
  2. Quantile Pathway: Nine separate regressor heads estimate quantiles (0.1 to 0.9). These are then refined using a rolling residual correction—a recursive adjustment based on the errors of the last 30 days to ensure the uncertainty intervals remain calibrated.
Figure 1: V1’s separate point and quantile pathways. The log-price point output is back-transformed before blending with the upper-price classifier. Rolling residual adjustments apply only to quantiles; all historical inputs precede the target date.

The authors evolved this system through several iterations (Exp5 and Exp7), adding features like:

  • Morning RTM Signals: Using the first few hours of the current day's real-time prices to predict the next day.
  • Auction Fundamentals: Incorporating "unmatched sell" volumes (surplus power that didn't clear).
  • Clear-Sky Solar Curves: A proxy for solar generation based on Global Horizontal Irradiance (GHI) at major solar hubs.

Datasets

The study utilizes historical price and volume data from the Indian Energy Exchange (IEX), specifically for the GDAM, DAM, and RTM segments.

  • Training/Evaluation Grid: The primary evaluation spans 257 delivery days (January 1 – September 14, 2026), totaling 24,672 blocks.
  • Holdout Set: A later, "harder" stretch from October 25 to November 16, 2026, was used to test the enhanced Exp5 and Exp7 variants.
  • Baselines: The models were compared against Chronos-2 (a pretrained neural time-series transformer) and simple persistence baselines (yesterday’s price).

Key Insights and Ideas

  • Regime Modeling Beats Pure Regression: Treating the price cap as a classification problem prevents the model from "averaging out" price spikes. Blending the classifier and regressor allows the model to transition smoothly between normal and capped regimes.
  • Domain Expertise Trumps "Zero-Shot" AI: Despite the hype around large pretrained time-series models, a well-tuned Gradient Boosted Decision Tree (GBDT) with domain-specific features (like the solar GHI curve) significantly outperformed the Chronos-2 transformer.
  • The Value of "Today's Morning": Incorporating RTM data from the morning of the auction (blocks 0–35) provided a powerful signal for the next day's prices, highlighting the strong temporal coupling between these markets.
  • Quantile Clock (Exp7): The authors found that error patterns vary by time of day. Exp7 uses a "clock mix" that weights different quantiles more heavily during specific windows (e.g., midday vs. evening peak) to improve accuracy.

Key Results

The "cap-aware" approach yielded substantial improvements across the board:

  • Aggregate Performance: Prod v1 achieved a Mean Absolute Error (MAE) of 397.9 Rs/MWh, a 31.8% reduction compared to Chronos-2 (583.3) and a 32.5% improvement over the previous day's profile.
  • Reliability: V1 maintained a Mean Absolute Percentage Error (MAPE) of 9.57%, whereas Chronos-2 sat at 13.75%.
  • Harder Horizons: In the late October/November evaluation, the enhanced Exp5 variant remained robust with an MAE of 716, while the Chronos-2 baseline's error ballooned to near 2,000.

Limitations and Future Work

  • Data Consistency: The authors noted that certain renewable bid fields became unavailable in the exchange's local files mid-way through 2026, limiting the features available for the v2 model.
  • Empirical Calibration: While the rolling residual correction improved coverage, it lacks a formal mathematical guarantee (like conformal prediction) for future days.
  • Financial Impact: The study focuses on forecast accuracy (MAE/MAPE) but does not yet translate these gains into actual trading profit and loss (P&L) or optimized bidding strategies.
  • Geographic Scope: The model is currently specific to the Indian GDAM; testing its generalizability to other capped markets (like parts of the EU or US) remains for future work.

Summary updated

Abstract

Accurate forecasting of Green Day-Ahead Market (GDAM) electricity prices is essential for market participants in the Indian power sector. We present a LightGBM-based forecasting framework for Indian Energy Exchange GDAM prices, which clear in 96 fifteen-minute intervals per day. The proposed system explicitly models price-cap behavior by separating capped-price events from normal market conditions. A classification head estimates the probability of the regulatory price cap (Rs 10,000/MWh), while a regression head predicts uncapped prices in logarithmic space. Final forecasts are obtained through probabilistic blending of these components and are accompanied by nine quantile forecasts for uncertainty estimation. Evaluated on 257 trading days between 1 January and 14 September 2026, the model achieves a mean absolute error (MAE) of 397.9 Rs/MWh and a mean absolute percentage error (MAPE) of 9.57%, substantially outperforming a Chronos-2 baseline (583.3 MAE) and a persistence baseline using the previous day's profile (589.1 MAE). We further introduce leak-safe variants that incorporate real-time market signals, unmatched sell volumes, and solar generation proxies. In a recent out-of-sample evaluation, the quantile-clock variant achieved the best short-term performance with an MAE of 789 Rs/MWh during 25-28 October, while the enhanced capped-price model achieved an MAE of 716 Rs/MWh across 21 evaluation days through 16 November. These results demonstrate the effectiveness of regime-aware probabilistic forecasting for high-volatility electricity markets.
BibTeX
@misc{pratap2026cap,
  author = {Shubham Pratap},
  title = {{Cap-Aware Boosting for Green Day-Ahead Electricity Price Forecasting}},
  year = {2026},
  eprint = {2611.21149},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.21149},
}
Figure