TDT4259 Group 69: Weekly reservoir inflow forecasting

Weekly inflow forecasts for Staavatn, Totak, Botnedalsvatn and Songavatn, evaluated on accuracy and on SINTEF's seasonal penalty.

pip install -r requirements.txt
python section3.py        # about 1 minute, writes everything to results/

File	What it does
`inflow.py`	Reusable helpers: load hourly data into Monday–Sunday weeks, features, penalty function, seasons
`section3.py`	Baselines, models, validation (1958–1989 / 1990–1999), test 2000–2014, tables and figures
`results/`	`section3_tables.xlsx` (all tables, one sheet each), `fig_*.png`, `summary.txt`

Method

Data and test setup: the forecast is issued on Monday for that week's mean inflow, using only data up to Sunday. A week's month and season follow its Thursday.
Validation: models are trained on 1958–1989 and checked on 1990–1999. All choices are made on the validation period.
Test: models are refitted every year on all earlier data, then forecast 2000–2014. The baseline is a fixed climatology from 1958–1999.
Models:
- Persistence
- Anomaly persistence
- Ridge regression
- Gradient boosting
- Penalty-aware: the most accurate model on validation (Ridge), with its forecast shifted up in Jan–Mar and down in Jul–Oct by a residual quantile q. q is chosen on validation as the largest value that keeps two guardrails:
  - |PBIAS| ≤ 25 % in every penalised season and reservoir (the "satisfactory" level in Moriasi et al., 2007)
  - MAE at most 10 % worse than Ridge
  The chosen value is q = 0.60.
Costly week: a week whose penalty is above the 90th percentile of the baseline's weekly penalties in 1990–1999. The threshold is set separately for each reservoir.
Bias guardrail (Table 1): |PBIAS| ≤ 10 % in the months without a penalty, for every reservoir (the "very good" level in Moriasi et al., 2007).
Flood weeks: the 10 % highest weekly inflows per reservoir. The threshold is set from 1958–1999.

For Section 3:
- KPI scorecard against Table 1
- validation results, test results, and test results per reservoir
- MAE and penalty by reservoir × season
- PBIAS
- flood weeks
- bootstrap confidence intervals, the two halves of the test period, the worst weeks
- runtime
Sheets starting with S4, for Section 4:
- the penalty converted to water (Mm³) and an approximate energy figure (GWh)
- the penalty share by reservoir
- errors larger than 5 % of reservoir storage
These are order-of-magnitude figures only. Water planned wrongly is not the same as water lost, and the energy factor is a system-wide average.