Network operators pay millions for imbalances. Solar and wind plants lose up to 15% of revenue due to inaccurate forecasts. Our AI-based renewable energy management system solves this: it predicts generation with MAPE below 8%, manages battery storage (BESS) through RL agents, and automates electricity market trading. Over 5+ years, we have completed 15+ projects for wind and solar farms totaling over 1 GW capacity. The core technology includes hybrid forecast models, LightGBM for correction, and a virtual power plant (VPP) for resource aggregation.
According to the NREL Solar Forecasting 2.0 study, pure physical models yield 10-15% error, while a hybrid with ML correction reduces it to 5-8%. Such approach enables effective participation in the electricity market and ancillary services. Results: 30-50% reduction in imbalance and up to 15% additional revenue.
Which AI models predict solar panel generation?
The hybrid model combines physics and machine learning.
Physical model
P_solar = G_poa × η × A × (1 - β × (T_module - T_ref)) G_poa — plane-of-array irradiance, η — efficiency, β — temperature coefficient. Irradiance is obtained from NWP models (ECMWF/GFS) by converting GHI to POA considering tilt and orientation.
ML correction of residuals
LightGBM learns from cloud cover (Cloud Cover Index from EUMETSAT MSG satellite), aerosols (AOD), panel temperature, and degradation factor. Result: MAPE 5-8%.
| Method | MAPE (daily) | Required data |
|---|---|---|
| Physical model | 10-15% | GHI, temperature, panel specs |
| Pure ML (LSTM) | 8-12% | Historical time series, weather |
| Hybrid (physics + LightGBM) | 5-8% | Physics + ML residuals, satellite imagery |
Wind power forecasting
Wind turbine power ∝ V³ in operating range. 10% error in wind speed → 30% error in power. We ensemble NWP models (ICON, GFS, ECMWF) via Bayesian Model Averaging (BMA).
def wind_power_ml(wind_speed, wind_direction, temperature, air_density, turbulence_intensity): features = np.array([wind_speed, np.sin(np.deg2rad(wind_direction)), np.cos(np.deg2rad(wind_direction)), temperature, air_density, turbulence_intensity]) return power_curve_model.predict(features.reshape(1, -1))[0] Why is an RL agent better than a rule-based controller for BESS?
Rule-based controllers cannot adapt to market changes or degradation. An RL agent (PPO or SAC) works in a continuous action space: state — SOC, forecasts, prices, system operator signals; reward — revenue from arbitrage minus degradation cost and imbalance penalties. Result: up to 20% additional revenue.
| Approach | Adaptability | Extra revenue | Implementation complexity |
|---|---|---|---|
| Rule-based | Low | — | Low |
| RL agent | High | up to +20% | Medium |
BESS tasks:
- Energy arbitrage: charge at low price, discharge at high
- Peak shaving: reduce load peaks (lower grid capacity cost)
- Frequency regulation: FCR/aFRR
- Smoothing: smooth renewable intermittency
How we do it: case study for a 50 MW solar plant
For one solar plant, we developed a hybrid forecast and RL agent for BESS. The first two weeks: data collection and SCADA audit. Then physical model calibration and LightGBM training on historical residuals. Concurrently, we built a BESS simulator for RL agent training. After 6 weeks, the system was ready for an A/B test: two weeks of RL agent showing 12% more revenue compared to the rule-based controller. After integration with BMS and the market, we ran a one-month pilot. Result: 35% imbalance reduction, 8% additional revenue from arbitrage.
What is included in the work
- ML forecast models (documented and versioned in MLflow)
- Integration with SCADA and BMS (Modbus/IEC 61850)
- Monitoring and control dashboard
- Operator training
- 3 months of technical support
VPP and market participation
We aggregate rooftop panels, BESS, EVs (V2G), and heat pumps into a virtual power plant. Dispatching via MILP (Gurobi/CPLEX) in real time. Participation in Day-Ahead, Intraday, and reserve markets (FCR, aFRR, mFRR). We guarantee forecast accuracy and compliance with system operator requirements.
Timeline and cost
Forecast + basic BESS control — 6-8 weeks. Full VPP with RL and market trading — 5-7 months. Cost is calculated individually after site audit. Get a consultation on AI system implementation for your facility and evaluate your savings potential.
Order a two-week pilot forecast on your data — we will show real accuracy on your site. Contact us for an audit and assessment of your renewable energy potential.