Electricity Demand Forecasting at Distribution and Household Levels Using Explainable Causal Graph Neural Network (X-CGNN)

This repository contains the code for our paper:

Electricity demand forecasting at distribution and household levels using explainable causal graph neural network

✍ Authors: Amir Miraki, Pekka Parviainen, Reza Arghandeh
📚 Journal: Energy and AI, Volume 16 (2024)
🔓 Open Access: Published by Elsevier under CC BY 4.0 License

In this work, we introduce a novel eXplainable Causal Graph Neural Network (X-CGNN) framework for multivariate electricity demand forecasting at both the household and distribution levels. X-CGNN leverages causal inference, graph neural networks, and temporal convolutions to provide state-of-the-art accuracy while ensuring model interpretability.

This approach addresses key challenges in power demand forecasting including capturing variable interdependencies and offering actionable insights for smart grid operations such as load balancing, resource allocation, and grid reliability.

---

📌 Key Features

• ✅ Causal Graph Discovery: Learns directed acyclic graphs (DAGs) from time series data to uncover causal dependencies.
• ✅ Graph Neural Networks: Enhances spatial modeling using Gated GNNs driven by causality.
• ✅ Temporal Convolution: Captures periodic and long-range temporal patterns through dilated convolutions.
• ✅ Intrinsic & Post-hoc Explainability: Combines global causal structure and local XAI (e.g., SHAP, LIME) explanations.
• ✅ Multi-level Forecasting: Validated on household and distribution-level real-world electricity demand datasets.

---

🧠 Model Architecture

The X-CGNN framework is composed of the following modules:

• 1. Causal Graph Discovery Module (CGD): Extracts causality graphs from multivariate time series using continuous optimization (NOTEARS).
• 2. Temporal Convolution Module (TC): Employs multiple dilated convolutions to capture diverse periodic patterns (e.g., daily/weekly).
• 3. Graph Neural Network Module (GNN): Applies Gated GNNs using the learned causal graph to model variable interactions for accurate forecasting.

---

📊 Use Cases

• Household-Level Forecasting: Predicting hourly electricity/heat consumption using weather and historical load data.
• Distribution-Level Forecasting: Forecasting energy demand in multiple city zones using SCADA data and meteorological variables.

---

📈 Performance Highlights

• 🚀 Achieves state-of-the-art results in both short- and long-term forecasting tasks.
• 📉 Outperforms classic models (AR, LSTM, XGBoost), and deep learning baselines (STGCN, MTGNN, Autoformer).
• 💯 Delivers over 50% improvement in MAE/RMSE vs. LSTM for 2-hour predictions at household level.

---

💡 Explainability

• Global: Interpretable causal graphs showing meaningful sensor/variable relationships.
• Local: Feature/time-step importance via SHAP, LIME, Feature Ablation, etc.
• These insights guide optimal sequence length and support model trust in real-world deployments.

---

▶️ Usage

⚙️ Dependencies: This project uses PyTorch Geometric. Please follow the official installation instructions compatible with your CUDA and PyTorch setup.

🔧 Run X-CGNN on the Tetuan dataset

python X_CGNN.py