Transformer meets wcDTW to improve real-time battery bids: A new approach to scenario selection
Sujal Bhavsar, Vera Zaychik Moffitt, Justin Appleby
TL;DR
The paper tackles stochastic battery bidding in real-time energy markets, where selecting representative historical scenarios is crucial for robust optimization across multiple uncertain products. It introduces a hybrid framework that fuses Transformer-based forecasting with weighted constrained Dynamic Time Warping ($wcDTW$) to identify analogous historical days that preserve interdependencies and forecast coherence across products. A probabilistic Transformer outputs across quantiles, while $wcDTW$ computes a distance $D_c$ to select scenarios with probabilities proportional to $1/D_c$, using a windowed distance $wcDTW(i, j)$ and quantile distances $d(x_i, y_j^q)$. The approach, validated on PJM data for July 2023, yields a $10\%$ revenue uplift for Energy RT in intra-day bidding and improves the stoch metric, demonstrating improved scenario quality, spike capture, and overall profitability for real-time energy arbitrage.
Abstract
Stochastic battery bidding in real-time energy markets is a nuanced process, with its efficacy depending on the accuracy of forecasts and the representative scenarios chosen for optimization. In this paper, we introduce a pioneering methodology that amalgamates Transformer-based forecasting with weighted constrained Dynamic Time Warping (wcDTW) to refine scenario selection. Our approach harnesses the predictive capabilities of Transformers to foresee Energy prices, while wcDTW ensures the selection of pertinent historical scenarios by maintaining the coherence between multiple uncertain products. Through extensive simulations in the PJM market for July 2023, our method exhibited a 10% increase in revenue compared to the conventional method, highlighting its potential to revolutionize battery bidding strategies in real-time markets.