Gated-LoRA Boosts Electricity Price Forecasting in Data-Scarce Markets

Forecasting electricity prices is critical for market participants but is complicated by price volatility, market-specific characteristics, and strong ties to anticipated system conditions. Existing supervised forecasting methods typically rely heavily on extensive historical data specific to a given market, which limits their applicability in new or data-poor markets. To address this, a new market-information-aware adaptation framework has been proposed. This framework leverages the Chronos-2 time-series foundation model and adapts it for day-ahead electricity price forecasting using a source-domain Gated Low-Rank Adapter (LoRA). The approach involves constructing a multi-source market information (MSMI) interface that aligns 7-day price context with various pre-clearing variables like supply-demand, reserve, maintenance, and generator capacity. The Gated-LoRA, which updates only about 1% of model parameters, is then trained without requiring target-market labels. A key innovation is the gate mechanism, which scales the frozen source adapter based on reserve-tightness and operating-state signals. Evaluated using a leave-one-market-out protocol across four Chinese provincial day-ahead spot markets, the framework reduced average MAE/RMSE by 6.24%/7.99% compared to a zero-shot Chronos-2 and by 3.05%/3.52% over vanilla Source-LoRA. These results highlight the effectiveness of structured market inputs and state-dependent gated LoRA in enabling practical transfer learning for electricity price forecasting in data-constrained environments.