The Prophet model has been trained using data from three weather stations located at different areas in the Tenerife island. The results obtained indicate a better fit to the time series than the classical fitted models.
The historical series is characterised by a marked dispersion in the summer months, influenced by the trade winds and the orography.
The results obtained are shown below:
| Coefficient | Puerto de La Cruz | Las Galletas | Pajalillos |
|---|---|---|---|
| MAE | 0.49 | 0.51 | 0.52 |
| RMSE | 0.64 | 0.67 | 0.69 |
| NSE | 0.62 | 0.65 | 0.61 |
| R Squared | 0.62 | 0.65 | 0.62 |
Several tests were conducted using additional regressors but no improvement in the results was observed.
The mean absolute percentage error when simulating the period from 01/05/2022 to 26/05/2023 at daily scale and 7 days horizon is:
| Coefficient | Puerto de La Cruz | Las Galletas | Pajalillos |
|---|---|---|---|
| MAPE | 22.30 | 19.29 | 15.87 |
The difference between using the previous week evapotranspiration and Prophet for a week ahead forecasting is shown below. The forecasted period ranges from 05/2022 to 05/2023.
The graph shows that both methods have an accuracy error of less than 20% most of the year. The accuracy offered by Prophet in the months with the highest demand, compared to the previous week's method, stands out.
| Method | Coefficient | Puerto de La Cruz | Las Galletas | Pajalillos |
|---|---|---|---|---|
| Prophet | MAPE | 11.79 | 11.72 | 8.47 |
| Sum (mm) | 1011.57 | 1173.12 | 1278.71 | |
| Previous week | MAPE | 14.20 | 12.26 | 11.14 |
| Sum (mm) | 1047.08 | 1262.17 | 1338.71 | |
| Historical | Sum (mm) | 1045.64 | 1264.17 | 1348.87 |
The results obtained using the minimum tempearature and solar radiaton as regressors are shown below:
| Coefficient | Puerto de La Cruz | Las Galletas | Pajalillos |
|---|---|---|---|
| MAE | 0.46 | 0.47 | 0.50 |
| RMSE | 0.59 | 0.63 | 0.70 |
| NSE | 0.68 | 0.70 | 0.62 |
| R Squared | 0.68 | 0.70 | 0.62 |