Location: Beijing, China 

Application Title: Policy-aware AI forecasting for electricity, carbon, coal and certificate markets 

Team Members: Wenqiang (Executive Sponsor), Xiao Liu, Wei Xu, Enjun Wang, Shusen Guo and Jianliang Yang. 

The challenge 

Energy markets are among the most complex and fast-moving parts of the global economy. 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. 

Unlike many financial markets, energy markets are shaped as much by government decisions as by supply and demand. New emissions rules, price caps, trading mechanisms or grid directives often alter incentives before their effects appear in numerical market data. Traditional forecasting tools rely primarily on historical price and volume data. As a result, they struggle to respond when policy interventions drive sudden changes, leading to inaccurate forecasts, higher financial risk and less efficient power scheduling. 

For grid operators, generators and trading desks, these gaps can translate into increased compliance costs, missed revenue opportunities and difficulty aligning dispatch decisions with evolving regulatory requirements. 

The solution 

Forecasting that integrates policy and market data 

The National Institute of Clean-and-Low-Carbon Energy (NICE); part of China Energy Investment Corporation Limited, developed an AI-powered forecasting system designed to combine numerical market analysis with policy understanding. 

The system 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, linking this information to live market data such as electricity prices, carbon allowance trades and coal indices. 

This allows the model to assess how new rules, policy signals or regulatory adjustments are likely to influence market behaviour before those effects are fully reflected in price movements. 

How it works 

Linking text, time and transactions 

The system operates on on-premises GPU clusters and is used by analysts, traders and planning teams for short- and medium-term forecasting, carbon cost estimation and dispatch planning. 

At a technical level, the platform combines retrieval-augmented generation (RAG), temporal text vector databases and multivariate transformer decoders. Rather than treating policy documents as static context, each document is timestamped, verified and processed through a retrieve–filter–extract pipeline to ensure relevance and credibility. 

Training data includes: 

• Approximately 1,000 authoritative policy and technical documents 
• Around 3,000 time-sensitive reports from financial and regulatory sources 
• 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, reducing training energy use by 95% compared with full retraining. 

How it differs from conventional approaches 

Dynamic alignment between policy and prices 

Conventional time-series forecasting models treat text and numerical data as loosely connected inputs, often aligned only during initial training. NICE’s system instead performs dynamic temporal retrieval, continuously sourcing and aligning new policy and market information as it emerges. 

A bidirectional alignment mechanism allows regulatory insights to inform numerical predictions and, in turn, allows market behaviour to update how policy language is interpreted. This creates ongoing synchronization between language and data. 

As a result, forecasts are not only more responsive but also explainable. Each prediction can be traced back to the specific policy announcement, regulatory change or market event that influenced it. 

Impact 

• Price forecasting accuracy improved, with error rates reduced to 0.69% MAPE for weekly electricity forecasts and 2.68% for daily forecasts 
• Carbon price forecasting accuracy of 1.21%, and 1% for coal markets 
• 50% reduction in price forecast error compared with existing methods 
• 95% reduction in GPU energy use through LoRA-based fine-tuning 
• 1% increase in forecast-driven trading optimisation
• Approximately USD 13.7 million in additional annual trading revenue 
• 3–5% reduction in carbon cost exposure through improved compliance alignment 
• Deployed across 300 internal users in four provinces 

Why it matters 

Improving decision-making in policy-driven markets 

Energy markets increasingly operate at the intersection of economics and regulation. NICE’s approach shows how AI systems that understand both policy language and numerical data can support more accurate forecasting and more transparent decision-making. 

By linking forecasts directly to the policy and market signals that drive them, the system helps analysts and planners respond earlier to change, reduce exposure to regulatory risk and align trading and dispatch decisions with evolving low-carbon objectives.