Five years of operation, in context
Every chart below is computed from 592,806 cleaned five-minute records. Each section ends with what the result means for the people who run, manage and regulate the plant.
Key findings
Generated automatically from the analysis results in plain language.
Generation and capacity factor
Annual energy compared with the design value of 317.5 MWh per year, and the monthly pattern behind it.
Annual energy and capacity factor
Monthly energy (MWh)
Daily energy with outages
Seasonality and water availability
The plant takes water from an agricultural channel, so output follows both rainfall and the irrigation calendar.
Running output by month
Power-duration curves
Physics-informed virtual sensing
The flow and efficiency channels on the PLC report zero. Penstock friction makes the measured head fall as flow rises, so flow and efficiency can be estimated from head and power.
Head falls as output rises
Estimated flow and efficiency by load
Outages and lost revenue
Lost energy is the normal running output for the calendar month multiplied by the time offline, valued at the 34 JPY/kWh feed-in tariff.
Largest outages by lost energy
Outage catalog
Bearing condition
A normal-behaviour model trained on 2021-22 predicts bearing 1 temperature from load and bearing 2. The gap between actual and predicted temperature is used as a health indicator.
Monthly bearing temperatures and elevated-regime share
Anomalies and alarm management
An Isolation Forest scores how unusual each running record is. The alert rules are tested against all five years to check that staff are not flooded with messages.
Isolation Forest flag rate by year
Raw threshold crossings vs managed notifications
Forecasting
Models were trained on 2021-2025 and tested on 2026, which they never saw. Persistence (tomorrow equals today) is the benchmark to beat.
Accuracy on 2026 test data
Daily energy: forecast vs actual, 2026
How the signals move together
Spearman rank correlation between signals while the unit is running.