🤖 machine learning

Market-Information-Aware Gated-LoRA of Foundation Models for Transferable Day-Ahead Electricity Price Forecasting

This paper proposes a market-information-aware gated-LoRA framework that adapts the Chronos-2 foundation model for transferable day-ahead electricity price forecasting in data-scarce markets by integrating multi-source market information and state-dependent gating, achieving significant error reductions over zero-shot and vanilla fine-tuning baselines.

Original authors: Hang Fan, Wei Wei, Shengwei Mei

Published 2026-08-13

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Hang Fan, Wei Wei, Shengwei Mei

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine trying to predict the weather in a brand-new city where you've never lived, using only a generic weather app designed for the whole world. You know the basics: it gets hot in summer and cold in winter. But you don't know that this specific city has a weird micro-climate caused by a nearby lake, or that the local wind patterns change every Tuesday. In the world of science, this is the challenge of "transfer learning." It's the art of taking a smart, pre-trained brain (a foundation model) that knows a lot about the general world and teaching it to understand a specific, tricky new situation without needing to relearn everything from scratch. Usually, these models are great at spotting patterns in numbers, like stock prices or temperature, but they struggle when the rules of the game change suddenly, like in a brand-new electricity market where prices can spike wildly.

Now, picture electricity prices not just as numbers on a screen, but as the result of a massive, high-speed auction. Every 15 minutes, power plants, wind farms, and solar panels shout out how much energy they can sell, while cities and factories shout out how much they need. The price is the "clearing price" that balances this supply and demand. The problem is that these markets are volatile and unique to every region. If a new market opens up, there's no history to study, making it a nightmare for traditional computers to predict prices accurately. This paper dives into that exact problem: how do we use a super-smart, pre-trained AI to predict electricity prices in a new market before we even have enough local data to teach it?

The researchers behind this study decided to test a new strategy using a powerful AI called Chronos-2, which is like a "universal time-traveler" trained on millions of different time-series patterns. Instead of just feeding it past prices, they gave it a "reference sheet" of the future. Think of it like this: if you were guessing the price of a concert ticket, you wouldn't just look at last year's prices; you'd also look at the band's popularity, the stadium's capacity, and whether it's raining that day. Similarly, the team built a special interface called the Multi-Source Market Information (MSMI) interface. This interface feeds the AI a 7-day history of prices, but crucially, it also hands it the known future details of the electricity market: how much wind and solar power will be available, how much electricity is needed, how many generators are under maintenance, and how much backup power (reserves) is ready.

The paper finds that simply giving the AI this rich "reference sheet" makes a huge difference. When they tested the AI with just basic info (like a standard weather app), it made mistakes. But when they gave it the full MSMI reference sheet, the errors dropped significantly. The average error (MAE) for predicting prices across four different Chinese provincial markets fell from 86.25 to 79.60 just by improving the information the AI received. This suggests that for electricity markets, the type of information you feed the model is just as important as the model itself.

But the researchers didn't stop there. They realized that even with the reference sheet, the AI needed a way to adjust its "personality" based on how tight or loose the market was feeling. Imagine a driver who knows the route perfectly but needs to switch driving styles depending on whether the road is icy or sunny. The team created a "Gated LoRA" mechanism. This is a tiny, lightweight add-on (updating only about 1% of the model's parameters) that acts like a smart volume knob. It looks at the current market state—how tight the power reserves are, how much renewable energy is flowing, and how volatile prices have been recently—and then tells the AI, "Hey, the market is tight today, so be more careful with your predictions," or "The market is calm, so you can relax."

The results showed that this "Gated LoRA" approach was the winner. It reduced the average error even further, down to 74.63, beating both the basic AI and a version that just learned from other markets without the smart "volume knob." The paper explicitly rules out the idea that a simple, fixed adjustment (like a global scalar) or a random guess would work; the "gate" had to be specifically tuned to the market's state to be effective. However, the authors are careful to note that while this method improves the accuracy of the average prediction, it doesn't perfectly fix the "confidence intervals" (the range where the price might actually land), suggesting that while the method is a strong practical step, it's not a magic bullet for every type of risk.

In the end, the paper suggests a step-by-step recipe for new electricity markets: start with the smart AI and the rich "reference sheet" (MSMI) to get a decent prediction immediately. Then, if you have data from similar, older markets, use the "Source-LoRA" to fine-tune the AI. Finally, use the "Gated LoRA" to let the AI react to the specific mood of the market every single day. It's a way to bring a world-class expert into a new, data-scarce town and have them start working effectively almost immediately, without needing years of local experience first.

Technical Summary: Market-Information-Aware Gated-LoRA of Foundation Models for Transferable Day-Ahead Electricity Price Forecasting

Problem Statement
Accurate day-ahead electricity price forecasting is critical for market participants but remains challenging due to price volatility, market-specific rules, and the complex interplay of renewable fluctuations, network constraints, and strategic behavior. Existing supervised methods rely heavily on historical data specific to a target market, rendering them ineffective for newly established markets or those with scarce data. While time-series foundation models (e.g., Chronos-2) offer zero-shot transfer capabilities, they typically operate as univariate forecasters or weakly exogenous models, failing to leverage the rich, known future market information (e.g., supply-demand, reserves, maintenance schedules) available prior to market clearing. The core challenge is to develop a transfer framework that exploits this day-ahead market information to adapt a foundation model to a target market without using target-market training labels.

Methodology
The authors propose a transferable forecasting framework built upon the Chronos-2 time-series foundation model, structured around three key components:

Multi-Source Market Information (MSMI) Interface:
Instead of relying solely on price history, the framework constructs an interface that aligns a 7-day price context with a comprehensive set of pre-clearing variables. This includes load, renewable generation (wind/PV), reserve conditions, maintenance capacity, generator availability, and intertie schedules. This interface treats these variables not as auxiliary features but as part of the economic information set from which prices are formed.
Source-Domain Gated LoRA Adaptation:
The model employs a Low-Rank Adaptation (LoRA) strategy trained exclusively on source markets (excluding the target market).
- Source-LoRA: A lightweight adapter updates approximately 1% of the model parameters (specifically the self-attention projections and output patch layers) using data from related source markets. The pretrained backbone remains frozen.
- Gated Mechanism: To address the fact that market dynamics vary by operating state, a market-state gate is introduced. This gate constructs a state vector $z_t$ based on reserve tightness, net load, renewable share, capacity margin, and recent volatility. A learned function scales the LoRA update strength ( $\gamma_\psi(z_t)$ ) dynamically. Crucially, the gate is initialized using source-domain data and trained only on source-market rolling tasks, ensuring no leakage from target-market labels.
Evaluation Protocol:
The study utilizes a Leave-One-Market-Out (LOMO) protocol. In each fold, one provincial market is held out as the target, while the remaining markets serve as the source domain for adaptation. This strictly simulates a scenario where no local training labels are available for the target market.

Key Contributions

MSMI Interface Design: The paper demonstrates that exposing the foundation model to a rich, multi-source market information interface is the primary driver of performance improvement, significantly outperforming interfaces limited to core drivers (load, wind, PV) or hand-crafted engineered features.
State-Dependent Adapter Calibration: The introduction of a market-state-gated LoRA mechanism allows the source adapter to adapt its correction strength based on real-time operating conditions (e.g., reserve tightness). This provides a compact calibration layer without requiring a full retraining of the backbone or a separate market-specific model.
Target-Market-Free Transfer Evidence: The work establishes a practical transfer path for data-scarce markets. It shows that combining a market-aware interface with source-domain LoRA adaptation yields superior results compared to zero-shot inference or vanilla LoRA, all without utilizing target-market labels during training.

Experimental Results
Experiments were conducted on four Chinese provincial day-ahead spot markets (Guangdong, Liaoning, Shandong, Shanxi).

Performance Gains: Relative to a market-information-aware zero-shot Chronos-2 baseline, the proposed framework reduced average MAE by 6.24% and RMSE by 7.99%. Compared to a vanilla Source-LoRA approach (without the gate), it achieved an additional reduction of 3.05% in MAE and 3.52% in RMSE.
Ablation Insights:
- The MSMI interface alone reduced zero-shot MAE from 86.25 to 79.60, confirming the necessity of known future market information.
- The gated mechanism provided the best results (Average MAE: 74.63). Control experiments showed that a learned global scalar or random gate initialization could not reproduce these gains, validating the importance of state-dependent scaling.
- Probabilistic Forecasts: The gated LoRA improved the Continuous Ranked Probability Score (CRPS) but resulted in slightly lower Prediction Interval Coverage Probability (PICP) for the 80% interval, indicating that while point accuracy improved, interval calibration remains an open issue requiring further refinement (e.g., conformal prediction).
Progressive Adaptation: The study outlines a deployment spectrum: starting with zero-shot inference, moving to source-domain LoRA when related data is available, applying gated calibration, and optionally using few-shot target refinement once local history accumulates.

Significance and Claims
The paper claims that for day-ahead electricity price forecasting, the task interface is a core modeling decision rather than a secondary preprocessing step. The results suggest that foundation models for energy markets must be evaluated not just on their backbone architecture but on their ability to ingest structured market information and adapt via lightweight, state-aware mechanisms.

The authors modestly conclude that while the gated LoRA offers a practical and lightweight transfer path for data-scarce markets, the gains over standard Source-LoRA are moderate and most pronounced in volatile markets (e.g., Liaoning). They acknowledge limitations, including the assumption that all future market information is available (which may not hold if data is delayed), the lack of causal isolation for specific input variables, and the need for broader validation across more diverse markets before claiming universal transferability. The work positions itself as a step toward bridging the gap between generic time-series foundation models and the specific economic realities of electricity market clearing.

More like this