The system reduced error rates to 0.69% MAPE for weekly electricity forecasts and 2.68% for daily forecasts, with carbon price forecasting accuracy of 1.21% and 1% for coal markets, a 50% reduction in price forecast error compared with existing methods. It delivered a 1% increase in forecast-driven trading optimisation, approximately USD 13.7 million in additional annual trading revenue and a 3-5% reduction in carbon cost exposure through improved compliance alignment. LoRA-based fine-tuning cut GPU energy use by 95% compared with full retraining. It is deployed across 300 internal users in four provinces.
Electricity prices, carbon allowance values, coal indices and renewable energy certificates can shift within hours in response to policy announcements, dispatch instructions or regulatory changes. Traditional forecasting tools rely primarily on historical price and volume data, so they struggle when policy interventions drive sudden changes, leading to inaccurate forecasts, higher financial risk and less efficient power scheduling.
The National Institute of Clean-and-Low-Carbon Energy (NICE), part of China Energy Investment Corporation Limited, developed an AI-powered forecasting system that combines numerical market analysis with policy understanding. It pairs a large language model specialised in energy policy and market regulation with a multivariate time-series forecasting network. Through a retrieval-augmented architecture, it continuously scans licensed policy databases, regulatory bulletins and trading reports and links them to live market data such as electricity prices, carbon allowance trades and coal indices.
The platform combines retrieval-augmented generation (RAG), temporal text vector databases and multivariate transformer decoders, and runs on on-premises GPU clusters. Each document is timestamped, verified and processed through a retrieve-filter-extract pipeline. Training data includes approximately 1,000 authoritative policy and technical documents, around 3,000 time-sensitive reports from financial and regulatory sources, and more than 200,000 rows of transaction data across electricity, coal, carbon and green-certificate markets. The model uses LoRA-based fine-tuning to adapt to the energy domain.
The system performs dynamic temporal retrieval, and a bidirectional alignment mechanism lets regulatory insights inform numerical predictions while market behaviour updates how policy language is interpreted. Each prediction can be traced back to the specific policy announcement, regulatory change or market event that influenced it. Analysts, traders and planning teams use it for short- and medium-term forecasting, carbon cost estimation and dispatch planning.
Details