RIS-LLM: A New Framework for Electricity Price Forecasting Integrating Reasoning Perception and Semantic Alignment

Core Insights: RIS-LLM (Reasoning Perception Semantic Large Language Model) bridges high-dimensional numerical time series with physical market drivers through three key innovations—reasoning feature extraction, time-series semantic alignment, and an LLM reasoning framework—providing fresh insights to the field of electricity price forecasting. Source Information: Original author Huang-Mg, published on GitHub (link: https://github.com/Huang-Mg/RIS-LLM) on May 24, 2026. Keywords: Electricity price forecasting, time series, large language model, causal reasoning, semantic alignment, energy AI, LLaMA, interpretable AI.

Electricity Price Forecasting (EPF) is a core indicator for modern energy system operations. However, traditional models struggle to capture multi-scenario influencing factors (supply-demand relationships, renewable energy output, fossil fuel prices, grid congestion, sudden weather changes, etc.) in high-volatility markets. Their nonlinear and multi-scale characteristics make them ineffective in responding to extreme events. Large Language Models (LLMs) excel at natural language reasoning and semantic understanding, but applying them to time series forecasting—especially bridging high-dimensional numerical data with physical market drivers—is a current challenge.

RIS-LLM's three-stage innovative framework:

1. Reasoning Feature Extraction (RFE) Module: Uses volatility-aware adaptive LOESS decomposition to separate transient fluctuations from structural trends. Steps include rolling volatility estimation, adaptive LOESS smoothing, Important Data Point (IDP) detection, and morphological feature extraction (5-dimensional features: amplitude, duration, slope, area, etc.).
2. Time Series Semantics (TSS) Alignment Module: Introduces a causal bias attention mechanism (builds a prior matrix based on Granger causality test) and generates interpretable semantic labels (e.g., "demand surge", "renewable energy surplus") via K-Means clustering.
3. LLM Reasoning Framework: Reprograms the frozen LLaMA-7B model through non-overlapping blocking, prototype selection (cosine similarity), four-component structured prompts (DataDesc+TaskInst+PatchInfo+PatchLabels), and cross-attention fusion. The entire process uses bfloat16 precision.

Data flow process: Input price sequence → RevIN normalization → RFE module (output morphological features M_feat) → TSS module (output attention features M_attn and semantic labels) → LLM reasoning framework (blocking + prototype selection + structured prompt + cross-attention) → frozen LLaMA-7B → output projection → denormalization → prediction results.

Technical Implementation: Modular design with core code modules including risllm/model.py (full pipeline), rfe.py (RFE module), tss.py (TSS module), granger.py (Granger causality precomputation), causal_attention.py (causal bias attention), etc. Also provides Time-LLM baseline comparison. Environment Requirements: Python3.9+, PyTorch2.0+, Transformers4.30+, CUDA GPU (24GB+ VRAM recommended for LLaMA-7B 4-bit quantization). Deployment: Place LLaMA-7B HuggingFace format weights (config.json, tokenizer.model, pytorch_model) in the llama/ directory.

The core contribution of RIS-LLM lies in the first systematic introduction of "reasoning perception" and "semantic alignment" into the field of time series forecasting. It converts numerical data into causal structural semantic representations, improving prediction accuracy while enhancing interpretability (predictions can be traced to specific physical drivers). Application Prospects: Valuable for scenarios like energy trading, grid dispatch, and energy storage optimization. It meets the demand for increased electricity price volatility in the global energy transition and represents a key technical direction of "interpretable AI + domain knowledge fusion".

Key takeaways summary:

• RFE module enables volatility-aware feature extraction;
• TSS module introduces Granger causality-prior attention mechanism;
• Semantic annotation converts numerical fluctuations into interpretable market event labels;
• Four-component structured prompts inject time-series information into frozen LLMs;
• The entire process uses bfloat16 precision to ensure efficiency;
• Modular design facilitates extension and reproduction.