Short-Term Load Forecasting (STLF) Using Graph Convolutional Recurrent Neural Networks (GCRNN)

This is the GitHub repository for the paper "Short-Term Load Forecasting Using Graph Convolutional Recurrent Neural Networks" EAAI 2022.

Brief Model Description

In this paper, we present a Graph Convolutional Recurrent Neural Network (GCRNN), based on the integration of Graph Convolutional Networks (GCN) and Long Short-Term Memory (LSTM) networks, to simultaneously incorporate the spatial and temporal information of similar users for more accurate short-term predictions. In this paper, we integrate GNN and LSTM networks into a unified network to extract the relational and temporal information simultaneously. In the proposed model, the LSTM’s conventional convolution gates are replaced with graph convolution networks. Unfolding the resulting network can capture the temporal behavior of a node and the effects that its neighbors impose through time. Tailored to the short-term load forecasting problem, the graph’s nodes represent various consumers, and the links between the nodes indicate a similarity in their corresponding consumption patterns. Such a model mitigates the issue of former studies in neglecting the relational information between the users. Moreover, it produces powerful and informative representations suitable for load forecasting without unnecessarily increasing the dimensionality of the data.

Structure of the Code and How to Use the Repository

The repository contains several important files and directories. The python files in the main directory are general functionalities common to both Low Carbon London (LCL) and Customer Behavior Trials (CBTs) datasets.

• layers.py - contains the implementation for a Graph Convolutional LSTM (GCLSTM) cell which is an LSTM cell which its gates has been replaced by graph convolutional layers.
• models.py - contains the unified GCRNN network followed by an MLP to be used for short-term load forecasting.
• utils.py - contains the helper functions used throughout the project.
• transforms.py - contains simple callables to transform numpy instances to Torch tensor instances.
• dataloader.py - contains the standard Torch DataLoader overrides for properly reading the LCL and CBT datasets.

Both LowCarbonLondon and CustomerBehaviorTrials directories contain similar python files for manipulation and experimentation on the corresponding datasets. Each directory contains the following files:

• data_inspection.py - contains necessary files to obtain preliminary information about the data.
• preprocessing.py - contains necessary files to transform and convert the data into the proper format.
• results_reader.py - contains a function for preparing and displaying the summary of the performance of a specified model.
• train_gcrnn.py - contains the necessary procedure for training the GCRNN model.
• train_lstm.py - contains the necessary procedure for training the LSTM model.
• train_svr.py - contains the necessary procedure for training the SVR model.
• train_ffnn.py - contains the necessary procedure for training the Feed Forward MLP model.
• train_arima.py - contains the necessary procedure for training the ARIMA model.
• test_gcrnn.py - contains the necessary procedure for testing and evaluating the GCRNN model and drawing informative and qualitative plots and drawings.

The dataset directory contains the datasets. Here we have not included the Low Carbon London dataset due to its volume. However, one can download the dataset from Kaggle. Once you have downloaded the dataset, extract it in a subdirectory named London in the dataset directory. Also, to respect the ownership rights we do not include the data files for the Customer Behavior Dataset. One can request the data from ISSDA and extract the contents in a subdirectory named CER in the dataset directory. These namings for subdirectories are essential for the right execution of the code.

To run experiments on the LCL dataset:

1. Run the generate_weather_dataset() function in preprocessing.py
2. Run the make_block_dataset() function in preprocessing.py
3. Run train_gcrn.py, train_lstm.py, train_ffnn.py, train_svr.py, or train_arima.py to run the experiment using one of the models.

To run experiments on the CBT dataset:

1. Run the group_maker() function in data_inspection.py
2. Run the prepare_data() function in preprocessing.py
3. Run train_gcrn.py, train_lstm.py, train_ffnn.py, train_svr.py, or train_arima.py to run the experiment using one of the models.