Utilize a dataset on electricity consumption and train a regression model to forecast future consumption.
The obtained .csv file contains columns such as Datetime, Temperature, Humidity, WindSpeed, GeneralDiffuseFlows, DiffuseFlows, PowerConsumption_Zone1, PowerConsumption_Zone2, PowerConsumption_Zone3. The average value was calculated across all three zones. A small window with a 10-minute interval was transformed into a window with a 1-hour interval, and the average values for each hour were calculated. Additionally, Datetime was split into columns 'Hour', 'Day', 'Month' as numerical features for training a regression model. The data were normolized & divided into test and train samples for model training and performance evaluation.
Histograms depicting the distribution of each feature were generated. Also power consumption graphics. Also, a graph illustrating the changes in power consumption over time was created.
Several regression models were trained, including SVR, LinearRegression, LassoRegression, RandomForestRegressor, GradientBoostingRegressor, AdaBoostRegressor, and DecisionTreeRegressor. Metrics such as MSE and r2-score were computed for each model. The best-performing models were GradientBoostingRegressor (r2-score = 0.96) and DecisionTreeRegressor (r2-score = 0.92). The complete table of metrics:
| Model | MSE | R2-score |
|---|---|---|
| SVR | 0.52 | |
| LinearRegression | 0.63 | |
| LassoRegression | 0.63 | |
| RandomForestRegressor | 0.86 | |
| GradientBoostingRegressor | 0.96 | |
| AdaBoostRegressor | 0.85 | |
| DecisionTreeRegressor | 0.92 |
Feature importances for best model (GradientBoostingRegressor) are:
- Temperature importances = 0.176;
- Humidity importances = 0.003;
- WindSpeed importances = 0.013;
- GeneralDiffuseFlows importances = 0.021;
- DiffuseFlows importances = 0.005;
- Hour importances = 0.697;
- Day importances = 0.009;
- Month importances = 0.076.