In this work, we introduce a data-centric framework, ESDF, to address the critical challenge of electricity load forecasting during extreme heatwaves and coldwaves. The core of our approach is a three-stage pipeline that sequentially addresses data synthesis, representation purification, and final prediction to overcome the fundamental limitations of data scarcity and the fidelity gap in synthetic data.
This repository includes codes for article "Electricity Load Forecasting Under Extreme Heat and Clod Waves" and a brief introduction of how to use them.
Authors: Nan Lu, Dalin Qin, Yangze Zhou, Lei Bai, Qingsong Wen, Chongqing Kang, Yi Wang
We propose an Extreme Synthesis-Disentanglement Forecasting (ESDF) framework to overcome the dual challenges of data scarcity and low-fidelity synthesis for load forecasting under heat and cold waves. The framework is designed as a three-stage pipeline that first strategically augments the training data with high-quality synthetic samples under extreme scenarios and then purifies the learned representations before making a final, resilient forecast.
The overview of the proposed framework is presented in the following figure, comprising three critical components: (1) Extreme Classifier-Guided Diffusion Synthesis, (2) Fidelity-Aware Representation Disentanglement, and (3) Resilience-Enhanced Forecasting.
Processed data can be found in the following OneDrive Link. Users can download the filefold "Data" from this link, and replace the filefold Data in this repository. Note that due to privacy concerns regarding certain data, we are only releasing two processed public datasets: Europe and PJM data. Each data file includes the hourly time index, electricity load, and temperature data.
Source Data:
Required libraries are provided in file requirements.txt. Please use "pip install -r requirements.txt" to ensure that codes can be executed.
Codes related to experiments are provided in filefold forecasting_using_generated_samples.py.
Codes for (1) Extreme Classifier-Guided Diffusion Synthesis can be found in the following files: Dataset_Loader_2D.py, diff_Model_2D. py, Models_2D.py, diff_training_2D.py, and generate_new_samples_2D.py; The trained diffusion model parameters will be saved in filefold Model_parameters.
codes for (2) Fidelity-Aware Representation Disentanglement and (3) Resilience-Enhanced Forecasting can be found in the following files: Forecasting_Models.py and Forecasting_model_training. The experimental results (nMAE and nRMSE of the tested regions) will be saved in a csv file (e.g. "Europe.csv"). Best forecasting models on the validation set will be saved as pt files.
Users can simply run the file Forecasting_model_training to execute the experiments. The experiments include some baselines and ablation comparison methods (see in Forecasting_model_training):
results_visualization(model_type='ANN', 'LSTM', or 'CNN'): simple baseline models
results_visualization(model_type='ANN', 'LSTM', or 'CNN', data_aug=1): simple baseline models with naive data augmentation strategies
results_visualization_nbeats(): NBEATS
results_visualization_impactnet(): ImpactNet
results_visualization_informer(): Informer
results_visualization_Autoformer(): Autoformer
results_visualization_dsn(): Proposed (integrated with different backbones)
Here we present the performance of all comparison methods on the selected datasets in terms of normalized mean absolute error (nMAE) and normalized root mean square error (nRMSE).
Codes and data related to figures in our manuscript and supplementary are provided in filefold Visualization. Users can run functions in Visualization/manuscript_fig.py to convert experimental results into visual images, which will be saved in Visualization/figures. Here are some examples of figures generated by our codes.
ESDF Systematically Outperforms Baseline Models Across All Datasets and Conditions. (a) -(b) Rose diagrams of average nMAE and nRMSE showing our framework achieves the lowest aggregate error in all six regions compared to eight baseline models. (c)-(f) Per-dataset error comparison under heatwaves (c, d) and coldwaves (e, f), plotting baseline model error (y-axis) against ESDF error (x-axis). The dense clustering of points above the unit line (
Representation Disentanglement Turns Unreliable Data Augmentation into a Robust Performance Enhancer. This figure shows that our disentanglement framework (ESDF) is critical for effectively using synthetic data. (a)-(b) Aggregate performance metrics show that naive data augmentation (blue bars) is a high-variance strategy with unpredictable, non-monotonic results. In contrast, ESDF (orange bar) delivers a substantial and consistent performance improvement. (c)-(d) Per-dataset scatter plots confirm this finding at a granular level. (e)-(f) Time-series plots reveal the practical impact: ESDF's predictions (red line) track the ground truth far more accurately during critical peak load periods. (g)-(h) Convergence curves demonstrate that ESDF enables the model to continuously achieve a lower error.
According to our experiments, some variables in codes can be changed if needed:
(1) the proportion of synthetic samples used: proportion
(2) weights of similarity and orhogonality penalties: alpha, beta
(3) the weight of the classifier guidance: numda