Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

A Decision-Focused Predict-then-Bid Framework for Strategic Energy Storage

Ming Yi, Yiqian Wu, Saud Alghumayjan, James Anderson, Bolun Xu

TL;DR

The paper tackles the problem of designing profitable energy storage arbitrage bids by bridging price prediction, bid design, and market clearing within a decision-focused framework. It introduces a tri-layer pipeline that forecasts prices, designs bid curves based on the storage’s opportunity value, and simulates market clearance in a differentiable manner using the implicit function theorem and a perturbed loss to enable end-to-end training. Key contributions include (1) a price-prediction–driven bid design anchored to the SoC dual variable, (2) a differentiable layer for the storage optimization via KKT-based gradients, and (3) a perturbation-based decision-focused loss that ensures smooth backpropagation through the market-clearing layer. Empirical results on NYISO data demonstrate significant profit gains over state-of-the-art benchmarks in both price-taker and price-maker settings, validating the framework’s practical impact for utility-scale storage bidding.

Abstract

This paper introduces a novel decision-focused framework for energy storage arbitrage bidding. Inspired by the bidding process for energy storage in electricity markets, we propose a predict-then-bid end-to-end method incorporating the storage arbitrage optimization and market clearing models. This is achieved through a tri-layer framework that combines a price prediction layer with a two-stage optimization problem: an energy storage optimization layer and a market-clearing optimization layer. We leverage the implicit function theorem for gradient computation in the first optimization layer and incorporate a perturbation-based approach into the decision-focused loss function to ensure differentiability in the market-clearing layer. Numerical experiments using electricity market data from New York demonstrate that our bidding design substantially outperforms existing methods, achieving the highest profits and showcasing the effectiveness of the proposed approach.

A Decision-Focused Predict-then-Bid Framework for Strategic Energy Storage

TL;DR

The paper tackles the problem of designing profitable energy storage arbitrage bids by bridging price prediction, bid design, and market clearing within a decision-focused framework. It introduces a tri-layer pipeline that forecasts prices, designs bid curves based on the storage’s opportunity value, and simulates market clearance in a differentiable manner using the implicit function theorem and a perturbed loss to enable end-to-end training. Key contributions include (1) a price-prediction–driven bid design anchored to the SoC dual variable, (2) a differentiable layer for the storage optimization via KKT-based gradients, and (3) a perturbation-based decision-focused loss that ensures smooth backpropagation through the market-clearing layer. Empirical results on NYISO data demonstrate significant profit gains over state-of-the-art benchmarks in both price-taker and price-maker settings, validating the framework’s practical impact for utility-scale storage bidding.

Abstract

This paper introduces a novel decision-focused framework for energy storage arbitrage bidding. Inspired by the bidding process for energy storage in electricity markets, we propose a predict-then-bid end-to-end method incorporating the storage arbitrage optimization and market clearing models. This is achieved through a tri-layer framework that combines a price prediction layer with a two-stage optimization problem: an energy storage optimization layer and a market-clearing optimization layer. We leverage the implicit function theorem for gradient computation in the first optimization layer and incorporate a perturbation-based approach into the decision-focused loss function to ensure differentiability in the market-clearing layer. Numerical experiments using electricity market data from New York demonstrate that our bidding design substantially outperforms existing methods, achieving the highest profits and showcasing the effectiveness of the proposed approach.
Paper Structure (20 sections, 3 theorems, 40 equations, 5 figures, 1 algorithm)

This paper contains 20 sections, 3 theorems, 40 equations, 5 figures, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Learning-Aided Energy Storage Operation
  4. Decision-Focused Learning
  5. Problem Formulation
  6. Bi-Level Storage Bidding Model
  7. Price-Taker and Price-Maker Reformulation
  8. Problem Statement and Learning Pipeline
  9. Methodology
  10. Bidding Design Based on Opportunity Value
  11. Differentiation of Energy Storage Optimization Layer
  12. Perturbed Decision-Focused Loss Function
  13. Algorithm Implementation
  14. Experiments
  15. Energy Storage Bidding Under Price-Taker Setting
  16. ...and 5 more sections

Key Result

Proposition 1

Marginal cost of the opportunity value function. Consider the predict-then-bid energy storage arbitrage problem given in eq:1a–eq:1e, assuming that both the primal and dual problems have non-empty feasible sets. The marginal cost of the opportunity value function $V_{t+1}$ with respect to the state The marginal costs with respect to discharge $p_t$ and charge $b_t$ at time $t$ are where $\bar{\t

Figures (5)

  • Figure 1: The bi-level framework illustrates the upper-level energy storage arbitrage problem and the lower-level market clearing process.
  • Figure 2: The proposed pipeline begins with a neural network-based predictor that forecasts electricity prices from the input features. The optimization layer then leverages the predicted prices, treated as a hidden reward, to solve an optimization problem and compute the dual variable of the state of charge (SoC). This dual variable is linearly transformed into charge/discharge bids. A subsequent market-clearing optimization layer processes these bids to generate charge/discharge dispatches. The algorithm employs backpropagation through both optimization layers to update the predictor's weights.
  • Figure 3: Annual cumulative profit comparison for the price-taker case, 5-minute resolution real-time price data in New York City. The energy storage model considers (a) a linear cost term, and (b) a combination of linear and quadratic cost terms.
  • Figure 4: Annual cumulative profit comparison for the price-taker case. The energy storage model is a linear cost term for (a) Long Island and (b) Western NY in NYISO.
  • Figure 5: Annual cumulative profit comparison for the price-maker case, considering (a) a linear price sensitivity model and (b) a cubic price sensitivity model.

Theorems & Definitions (3)

  • Proposition 1
  • Proposition 2
  • Proposition 3