LSTM-MultiStep-Forecasting

Implementation of Electric Load Forecasting Based on LSTM (BiLSTM). Including direct-multi-output forecasting, single-step-scrolling forecasting, multi-model-single-step forecasting, multi-model-scrolling forecasting, multi task learning, and seq2seq forecasting.

Environment

pytorch==1.10.1+cu111

numpy==1.18.5

pandas==1.2.3

python==3.7.3

Tree

1. args.py is a parameter configuration file, where you can set model parameters and training parameters.
2. data_process.py is the data processing file. If you need to use your own data, then you can modify the load_data function in data_process.py.
3. Four models are defined in models.py, including LSTM, bidirectional LSTM, multi task learning LSTM, and seq2seq.
4. model_train.py defines the training functions of the models in the six multi-step prediction methods.
5. model_test.py defines the testing functions of the models in the six multi-step prediction methods.
6. The trained model is saved in the models folder, which can be used directly for testing. The mms folder saves the model of multi-model-scrolling forecasting, and the mmss folder saves the model of multi-model-single-step forecasting.
7. Data files in csv format are saved under the data file.

Usage

First switch the working path:

cd algorithms/

Then, execute in sequence:

python multi_model_scrolling.py --epochs 50 batch_size 30
python multi_model_single_step.py --epochs 50 batch_size 30
python multiple_outputs.py --epochs 50 batch_size 30
python seq2seq.py --epochs 50 batch_size 30
python single_step_scrolling.py --epochs 50 batch_size 30
python multi_task_learning.py --epochs 50 batch_size 30

If you need to change the parameters, please modify them manually in args.py.

Result

Predict the next 12 steps, epochs=50, bacth_size=30, and the results of the 5 methods are shown in the following table:

method	1	2	3	4	5
MAPE/%	9.33	10.62	9.94	22.45	9.09