Multi-Step Time Series Forecasting of Power Consumption: Evaluation of Various Models Including CNN, LSTM, CNN-LSTM, and Multiple Linear Regression
Objectives of the project:
This project has two specific aims:
• Predicting weekly energy usage for individual household unit by analyzing its historical consumption levels,
• Evaluating the strengths and weaknesses of various modeling approaches for time series forecasting.
On a broader scale, the objective includes optimizing energy forecasting to balance supply and demand effectively.
1. Import the necessary libraries
2. Import and explore the dataset
• df.shape, df.head(), df.tail(), df.info()
3. Data Pre-processing and Exploratory Data Analysis (EDA)
• 3.1 Feature engineering (combining the date and time columns)
• 3.2 Data type conversion (converting all the features to float64 data type)
• 3.3 Handle missing values
• 3.4 Problem framing (Downsample the dataset on daily basis)
• 3.5 Handle duplicated values
• 3.6 Outliers Detection
4. Train-Test split (to avoid data leakage)
5. Scale the data (MinMax Scaling)
6. Convert the data into a supervised regression problem
7. Build models using Deep Learning Architectures (CNN, LSTM, CNN-LSTM), make predictions, and perform evaluations on them
• 7.1 CNN Models
• 7.2 LSTM Models
• 7.3 CNN-LSTM Models
8. Build a model using Multiple Linear Regression (MLR), make predictions, and perform evaluations on it
9. Best Performance Metrics of Four Models at Project Conclusion
10. Future Work