CPU Forecasting — Journal Experiments

Multi-model CPU usage forecasting framework. Trains and evaluates deep learning models on Azure and Alibaba VM datasets across a 3-phase transfer learning protocol.

Models

Model	File
BiLSTM	bilstm/bilstm.py
LSTM	lstm/lstm.py
RNN	rnn/rnn.py
Temporal Fusion Transformer (TFT)	tft/tft.py
Informer	informer/informer.py
ARIMA (baseline)	arima/arima.py

Experiment Protocol

For each model and each input/output configuration (6-1, 6-2, 6-3, 12-1, 12-2, 12-3, 24-1, 24-2, 24-3):

1. Phase 1 — Train on Azure → Evaluate on Azure test set
2. Phase 2 — Fine-tune on Alibaba → Evaluate on Alibaba test set
3. Phase 3 — Evaluate on Azure again (catastrophic forgetting check)

Results are saved to <model>_azure_results.txt, <model>_alibaba_results.txt, and <model>_azure_post_ft_results.txt.

Dataset

dataset/
├── azure/       # 100 Azure VM CPU time series (.csv)
└── alibaba/     # 100 Alibaba VM CPU time series (.csv)

Each CSV must contain an avg cpu column (or use the first column as fallback).

Usage

# Run all models
python main.py --model all

# Run specific models
python main.py --model bilstm rnn lstm

# Custom dataset paths and epochs
python main.py --model informer --azure_dir dataset/azure --alibaba_dir dataset/alibaba --num_epochs 30

# Override learning rates
python main.py --model tft --lr 0.0005 --ft_lr 0.00005

Arguments

Argument	Default	Description
--model	all	Model(s) to run: all, bilstm, rnn, lstm, informer, tft
--azure_dir	dataset/azure	Path to Azure CSV directory
--alibaba_dir	dataset/alibaba	Path to Alibaba CSV directory
--num_epochs	50	Max training epochs (early stopping applies)
--lr	per-model	Learning rate override for Phase 1 training
--ft_lr	per-model	Learning rate override for Phase 2 fine-tuning

ARIMA baseline

python arima/arima.py

Requirements

torch
numpy
pandas
scikit-learn

Project Structure

.
├── main.py              # Experiment runner
├── dataset/
│   ├── azure/
│   └── alibaba/
├── bilstm/bilstm.py
├── lstm/lstm.py
├── rnn/rnn.py
├── tft/tft.py
├── informer/informer.py
└── arima/arima.py