GridLDM: Unified Latent Diffusion for Cross-Domain, Language-Conditioned Power-System Time-Series Synthesis

April 14, 2026 · View on GitHub

Overview

GridLDM is a unified latent diffusion framework for controllable power-system time-series generation.
Instead of training a separate generator for each dataset and each conditioning schema, GridLDM learns a shared latent generative prior across heterogeneous power-system domains and conditions generation with natural-language prompts through cross-attention.

The core idea is:

  1. map variable-length time series into a compact common latent space with a prompt-conditioned autoencoder,
  2. train a latent diffusion model in that shared space,
  3. use natural-language prompts encoded by a pretrained language model to control conditional generation,
  4. decode synthetic latents back into the original time-series space.

This repository contains the training and sampling code for the GridLDM pipeline, together with several baseline models used for comparison.

GridLDM framework

Method Summary

GridLDM consists of two main stages:

1. Prompt-conditioned autoencoder

A Transformer-based conditional autoencoder maps variable-length time series into a fixed latent representation while injecting the text-conditioning signal into both the encoder and decoder.

2. Latent diffusion model

A 1D U-Net diffusion model is trained in latent space and uses cross-attention over text-token embeddings at each denoising step. During inference, classifier-free guidance is used to improve prompt alignment.

This design makes the model:

  • more stable than direct generation in raw signal space,
  • easier to reuse across domains with different lengths and temporal structures,
  • more flexible than schema-fixed label conditioning.

How to use GridLDM

The recommended workflow for reproducing the project is:

Step 0. Create environment and install dependencies

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Step 1. Process the prompts and build text embeddings

Use data_process.py to convert prompt JSONL files into token-level BERT embeddings.

python data_process.py

Step 2. Train the prompt-conditioned autoencoder

Train the multi-domain conditional autoencoder:

python encoder_decoder_conditional.py

This stage learns a shared latent representation across domains and saves:

cond_autoencoder.pt

Step 3. Train GridLDM (latent diffusion)

Train the latent diffusion model in the frozen autoencoder latent space:

python cross_diffusion_mix.py

This stage loads the pretrained autoencoder, encodes time series into latents, and trains the text-conditioned diffusion backbone. The default output checkpoint is:

gridldm_diffusion.pt

Step 4. Run sampling / generation

Generate synthetic time series from prompts with the unified sampling script:

python gridldm_sampling.py

You can also sample only selected domains:

python gridldm_sampling.py --domains wind solar
python gridldm_sampling.py --domains ev
python gridldm_sampling.py --domains commercial_load transient_voltage

Generated samples are saved under domain-specific output folders such as:

samples_wind/
samples_solar/
samples_ev/
samples_comm/
samples_trans/

Step 5. Fine-tune if needed

If you want to adapt GridLDM to a shifted prompt format or a new conditioning style, use the fine-tuning script:

python cross_diffusion_mix_finetune.py

This script freezes the autoencoder and fine-tunes only the diffusion backbone on a small subset of new prompt–data pairs.

Step 6. Train other baselines

The repository also includes several baseline training scripts:

  • Conditional VAE

    python cond_VAE.py

  • Conditional GAN

    python cond_GAN.py

  • Direct diffusion in raw signal space

    python direct_diffusion.py

These baselines are useful for ablation studies and comparison against GridLDM.

Data Format

Dataset

We evaluate GridLDM on five power-system time-series domains: transient voltage response, commercial user load, EV charging profiles, wind generation, and solar generation. The datasets span different temporal resolutions, operating contexts, and conditioning attributes, which makes them suitable for testing unified cross-domain generation.

  • Transient voltage response: short-term post-disturbance voltage trajectories collected from benchmark power networks.
  • Commercial user load: daily commercial energy usage traces at 15-minute resolution.
  • EV charging profiles: vehicle-level charging records aggregated into normalized 24-hour charging profiles under different charging-location allocations.
  • Wind generation: daily wind generation profiles organized by load zone across major U.S. ISOs from 2018 to 2020, paired with zone-level weather measurements.
  • Solar generation: daily solar generation profiles organized by load zone across major U.S. ISOs from 2018 to 2020, paired with zone-level weather measurements.

For the renewable domains, the original data are at 5-minute resolution and are downsampled to hourly resolution to form 24-hour daily profiles. In total, the full experimental setup contains approximately 130k prompt–data pairs.

Important note: the commercial load dataset is not included in this repository because it is confidential.

For additional details on dataset construction, conditioning attributes, and experiment setup, please refer to the paper.

Prompt Conditions

GridLDM supports multiple natural-language prompt styles per domain. prompts are stored as JSONL files under prompts/<domain>/... and are converted into token embeddings by data_process.py. The same underlying signal can be paired with different prompt formulations, which helps evaluate robustness to linguistic variation and conditional consistency.

For example:

  • EV prompts describe date, week type, load type, and allocation
  • wind / solar prompts describe zone, date, and weather / pattern statistics
  • transient prompts describe event type, bus, and network
  • commercial-load prompts describe user, date, and weekday/weekend

Citation

If you use this codebase in your research, please cite the GridLDM paper.

Contact

If you find this repository useful or have questions, please reach out at cyuting@tamu.edu.