Revenue comes from three directions
A VPP operator’s revenue today is roughly built from three sources: demand-response subsidies, ancillary-service compensation, and price arbitrage in the spot market. Each demands a different capability — demand response tests whether resources are reliably dispatchable, ancillary services test response speed, and spot arbitrage rests almost entirely on forecast accuracy.
Hence a half-joking line in the industry: in the end, a virtual power plant is an algorithms company.
How accuracy converts to money
Take PV power forecasting. Each percentage point gained in ultra-short-term (15 min – 4 h) accuracy directly reduces deviation penalties; improvements in day-ahead accuracy affect bidding strategy and the revenue ceiling in the day-ahead market. Combined, for a hundred-megawatt aggregated portfolio the annual difference can be substantial.
Load forecasting works the same way. Moving day-ahead accuracy from 85% to over 90% gives tariff optimization and demand management a reliable basis, and the resulting drop in basic electricity charges is close to deterministic.
No silver bullet in method
In practice we use an ensemble: numerical weather prediction supplies the physical prior, LSTM captures local temporal features, Transformer-family models handle long-range dependencies, and random forests model nonlinear residuals. Reinforcement learning and evolutionary algorithms then optimize storage dispatch, with full-cycle revenue maximization as the reward function.
None of this is mystical. The difficulty is engineering — data cleaning, gap filling, similar-day construction, online model updates, and integration with the customer’s existing SCADA / EMS. It is usually this unglamorous work that separates the field.
This column is industry commentary and technical opinion, and does not constitute investment advice.